Snapdragon is the future with Qualcomm's Leendert van Doorn

hi I'm Scott Hanselman this is another hi I'm Scott Hanselman this is another episode of Hansel minutes today I'm episode of Hansel minutes today I'm episode of Hansel minutes today I'm chatting with Lander vanor he's a senior chatting with Lander vanor he's a senior chatting with Lander vanor he's a senior vice president at Qualcomm previously he vice president at Qualcomm previously he vice president at Qualcomm previously he was a distinguished engineer at was a distinguished engineer at was a distinguished engineer at Microsoft's Azure core team before Microsoft's Azure core team before Microsoft's Azure core team before joining Microsoft he was a corporate joining Microsoft he was a corporate joining Microsoft he was a corporate fellow at AMD you've been doing some fellow at AMD you've been doing some fellow at AMD you've been doing some amazing work sir and now you're at the amazing work sir and now you're at the amazing work sir and now you're at the Forefront of AI at Qualcomm I'm thrilled Forefront of AI at Qualcomm I'm thrilled Forefront of AI at Qualcomm I'm thrilled to be chatting with you today thank you to be chatting with you today thank you to be chatting with you today thank you and likewise really looking forward to and likewise really looking forward to and likewise really looking forward to this podcast you know um I want to start this podcast you know um I want to start this podcast you know um I want to start with a little bit of historical context with a little bit of historical context with a little bit of historical context and tell me if this works many many many and tell me if this works many many many and tell me if this works many many many years ago when the IBM uh 486 and 386 years ago when the IBM uh 486 and 386 years ago when the IBM uh 486 and 386 came out there was an SX and a DX and came out there was an SX and a DX and came out there was an SX and a DX and they had a co-processor on the DX and they had a co-processor on the DX and they had a co-processor on the DX and they didn't on the SX and people would they didn't on the SX and people would they didn't on the SX and people would buy the one with the co-processor buy the one with the co-processor buy the one with the co-processor because they could do floating point because they could do floating point because they could do floating point they could do excel they could do cooler they could do excel they could do cooler they could do excel they could do cooler stuff and now we're seeing two amazing stuff and now we're seeing two amazing stuff and now we're seeing two amazing things that are happening in my opinion things that are happening in my opinion things that are happening in my opinion on the PC one we're seeing a move on the PC one we're seeing a move on the PC one we're seeing a move towards arm and we're seeing this towards arm and we're seeing this towards arm and we're seeing this additional npu co-processor it's like additional npu co-processor it's like additional npu co-processor it's like that's two innovations happening at the that's two innovations happening at the that's two innovations happening at the exact same time is that a reasonable exact same time is that a reasonable exact same time is that a reasonable analogy um it is uh so obviously the analogy um it is uh so obviously the analogy um it is uh so obviously the transition to R right is giving us a transition to R right is giving us a transition to R right is giving us a more modern Isa and in many ways a more more modern Isa and in many ways a more more modern Isa and in many ways a more efficient Isa and so if you look at what efficient Isa and so if you look at what efficient Isa and so if you look at what we did recently with the uh uh selite we we did recently with the uh uh selite we we did recently with the uh uh selite we give an give an give an enormous battery Lifetime right 19 hours

enormous battery Lifetime right 19 hours enormous battery Lifetime right 19 hours I think for the Dell XPS 26 hours for HP I think for the Dell XPS 26 hours for HP I think for the Dell XPS 26 hours for HP and so that's unheard of and that's all and so that's unheard of and that's all and so that's unheard of and that's all because of the efficiency of the because of the efficiency of the because of the efficiency of the processor and it's also in part the processor and it's also in part the processor and it's also in part the efficiency of the ISA then there is a efficiency of the ISA then there is a efficiency of the ISA then there is a whole other component of it as well whole other component of it as well whole other component of it as well because doing things on a general because doing things on a general because doing things on a general purpose computer still takes power right purpose computer still takes power right purpose computer still takes power right and then for dedicated tasks multimedia and then for dedicated tasks multimedia and then for dedicated tasks multimedia but also AI you want to have a dedicated but also AI you want to have a dedicated but also AI you want to have a dedicated accelerator that is just optimized for accelerator that is just optimized for accelerator that is just optimized for doing that uh like take for example doing that uh like take for example doing that uh like take for example playing a video right for playing a playing a video right for playing a playing a video right for playing a video you don't need to keep your more video you don't need to keep your more video you don't need to keep your more powerful uh and therefore more energy powerful uh and therefore more energy powerful uh and therefore more energy hungry main processor busy you can just hungry main processor busy you can just hungry main processor busy you can just keep a small accelerator on the site keep a small accelerator on the site keep a small accelerator on the site busy and it does all the playing and busy and it does all the playing and busy and it does all the playing and that's how you get to these really long that's how you get to these really long that's how you get to these really long battery lifetimes same is true for battery lifetimes same is true for battery lifetimes same is true for accelerator for for AI accelerators accelerator for for AI accelerators accelerator for for AI accelerators right if you look at recall continuously right if you look at recall continuously right if you look at recall continuously runs in the background and monitors runs in the background and monitors runs in the background and monitors things and learns about your environment things and learns about your environment things and learns about your environment uh by running that on a dedicated npu it uh by running that on a dedicated npu it uh by running that on a dedicated npu it takes about at 30 minutes of your takes about at 30 minutes of your takes about at 30 minutes of your lifetime of the battery lifetime of an lifetime of the battery lifetime of an lifetime of the battery lifetime of an entire run as supposed if you were to do entire run as supposed if you were to do entire run as supposed if you were to do this on a GPU where it would take 6 this on a GPU where it would take 6 this on a GPU where it would take 6 hours or more so yeah accelerators is hours or more so yeah accelerators is hours or more so yeah accelerators is all about how can you more efficiently all about how can you more efficiently all about how can you more efficiently do your compute by having a dedicated do your compute by having a dedicated do your compute by having a dedicated engine board one of the things I like to engine board one of the things I like to engine board one of the things I like to do on the podcast is expand three-letter do on the podcast is expand three-letter do on the podcast is expand three-letter acronyms uh tlas uh so you said Isa

acronyms uh tlas uh so you said Isa acronyms uh tlas uh so you said Isa instruction set architecture that our instruction set architecture that our instruction set architecture that our audience might not be familiar with why audience might not be familiar with why audience might not be familiar with why is that such an important thing that is that such an important thing that is that such an important thing that we're moving away from the x86 x64 we're moving away from the x86 x64 we're moving away from the x86 x64 instruction set architecture the 6 instruction set architecture the 6 instruction set architecture the 6 instruction sets architecture has a lot instruction sets architecture has a lot instruction sets architecture has a lot of Legacy right it it's like it's a 40y of Legacy right it it's like it's a 40y of Legacy right it it's like it's a 40y old uh essay and so we've learned a lot old uh essay and so we've learned a lot old uh essay and so we've learned a lot of things and so the arm instruction set of things and so the arm instruction set of things and so the arm instruction set had the benefit of starting much later had the benefit of starting much later had the benefit of starting much later so they could take a lot of those so they could take a lot of those so they could take a lot of those learnings from x86 and other instruction learnings from x86 and other instruction learnings from x86 and other instruction sets and then build a more optimized sets and then build a more optimized sets and then build a more optimized instruction set around it that takes instruction set around it that takes instruction set around it that takes less uh Power um and I noticed that less uh Power um and I noticed that less uh Power um and I noticed that Qualcomm and this this is kind of a uh Qualcomm and this this is kind of a uh Qualcomm and this this is kind of a uh my my analysis of the thing seems to my my analysis of the thing seems to my my analysis of the thing seems to have shifted to a developer first have shifted to a developer first have shifted to a developer first approach which I find very attractive as approach which I find very attractive as approach which I find very attractive as someone who's a developer first like you someone who's a developer first like you someone who's a developer first like you know who are we doing this for this is know who are we doing this for this is know who are we doing this for this is for the developer we want to make them for the developer we want to make them for the developer we want to make them happier because if you make the happier because if you make the happier because if you make the developers happier that's the you know developers happier that's the you know developers happier that's the you know the tip of the spear and everything else the tip of the spear and everything else the tip of the spear and everything else comes from there was that something you comes from there was that something you comes from there was that something you were involved in or was it like that were involved in or was it like that were involved in or was it like that when you arrived at qualcom no no very when you arrived at qualcom no no very when you arrived at qualcom no no very actively involv it so I joined qualcom actively involv it so I joined qualcom actively involv it so I joined qualcom two years ago two years ago two years ago um and I specifically joined to help um and I specifically joined to help um and I specifically joined to help with software strategy and with some of with software strategy and with some of with software strategy and with some of our ring data center initiatives uh and our ring data center initiatives uh and our ring data center initiatives uh and that was my background at Microsoft that was my background at Microsoft that was my background at Microsoft right and Azure as you just mentioned I right and Azure as you just mentioned I right and Azure as you just mentioned I was responsible for getting armed server was responsible for getting armed server was responsible for getting armed server into Azure and developing both windows into Azure and developing both windows into Azure and developing both windows but also the entire edure software stack

but also the entire edure software stack but also the entire edure software stack both storage and virtualization stack both storage and virtualization stack both storage and virtualization stack and moving that to uh to to arm and and moving that to uh to to arm and and moving that to uh to to arm and getting that stood up measure so when I getting that stood up measure so when I getting that stood up measure so when I came to qualcom one of my first came to qualcom one of my first came to qualcom one of my first observations was observations was observations was that we we we weren't very developer that we we we weren't very developer that we we we weren't very developer friendly now and it's all very friendly now and it's all very friendly now and it's all very explainable because we are historically explainable because we are historically explainable because we are historically a cell phone company that is very a cell phone company that is very a cell phone company that is very aggressively trying to diversify but aggressively trying to diversify but aggressively trying to diversify but we're still going through this journey we're still going through this journey we're still going through this journey and so one of the things with cell phone and so one of the things with cell phone and so one of the things with cell phone is you got a handful of customers that is you got a handful of customers that is you got a handful of customers that you work with we are also an IP company you work with we are also an IP company you work with we are also an IP company so therefore very protective of of of so therefore very protective of of of so therefore very protective of of of our knowledge right and so the company our knowledge right and so the company our knowledge right and so the company was set up for that and if I then looked was set up for that and if I then looked was set up for that and if I then looked at the uh our growth businesses XR PC at the uh our growth businesses XR PC at the uh our growth businesses XR PC iot Auto these all have a lot of iot Auto these all have a lot of iot Auto these all have a lot of different developers and so I made the different developers and so I made the different developers and so I made the argument that this what we call argument that this what we call argument that this what we call internally develop first is like hey we internally develop first is like hey we internally develop first is like hey we really need to enable developers to use really need to enable developers to use really need to enable developers to use our Hardware to use our platform able to our Hardware to use our platform able to our Hardware to use our platform able to develop for our platforms with and develop for our platforms with and develop for our platforms with and without our help so the with our help without our help so the with our help without our help so the with our help we're really good at it we've got a we're really good at it we've got a we're really good at it we've got a large customer engineering organization large customer engineering organization large customer engineering organization but it doesn't scale right it doesn't but it doesn't scale right it doesn't but it doesn't scale right it doesn't scale to thousands 10 thousands 100 scale to thousands 10 thousands 100 scale to thousands 10 thousands 100 thousand developers so we need to get thousand developers so we need to get thousand developers so we need to get this this this without our help or what this this this without our help or what this this this without our help or what I call no touch or low no low touch I call no touch or low no low touch I call no touch or low no low touch enablement and so that's something we've enablement and so that's something we've enablement and so that's something we've been very focused on as a result of it I been very focused on as a result of it I been very focused on as a result of it I I wrote this thing because every company

I wrote this thing because every company I wrote this thing because every company needs to go through that what I called needs to go through that what I called needs to go through that what I called the open Manifesto where I basically the open Manifesto where I basically the open Manifesto where I basically laid it out I said look our growth laid it out I said look our growth laid it out I said look our growth businesses are not going to reach its businesses are not going to reach its businesses are not going to reach its full potential if we don't work better full potential if we don't work better full potential if we don't work better with developers and give the developers with developers and give the developers with developers and give the developers what they want and what they need which what they want and what they need which what they want and what they need which is access to information right don't is access to information right don't is access to information right don't enumber this with license terms or enumber this with license terms or enumber this with license terms or things like that we need to be more open things like that we need to be more open things like that we need to be more open um and so that's the journey we've been um and so that's the journey we've been um and so that's the journey we've been actively on for the last two years and actively on for the last two years and actively on for the last two years and by the way all of this is understandable by the way all of this is understandable by the way all of this is understandable how we got here right qualcom got there how we got here right qualcom got there how we got here right qualcom got there uh and we are a very successful company uh and we are a very successful company uh and we are a very successful company right so it's there's nothing wrong with right so it's there's nothing wrong with right so it's there's nothing wrong with that approach it's just that in the new that approach it's just that in the new that approach it's just that in the new world where we have to address many many world where we have to address many many world where we have to address many many markets deal with many many different markets deal with many many different markets deal with many many different kind of developers we need a different kind of developers we need a different kind of developers we need a different approach and that's really what's comes approach and that's really what's comes approach and that's really what's comes down to what we refer to as developer down to what we refer to as developer down to what we refer to as developer first now there's a ton of silicon first now there's a ton of silicon first now there's a ton of silicon Innovation going on but you called out Innovation going on but you called out Innovation going on but you called out uh you're doing uh XR extended reality uh you're doing uh XR extended reality uh you're doing uh XR extended reality doing VR you're doing uh you know Linux doing VR you're doing uh you know Linux doing VR you're doing uh you know Linux on iot on the edge you've got these s on iot on the edge you've got these s on iot on the edge you've got these s so's these systems on a chip for every so's these systems on a chip for every so's these systems on a chip for every form factor and I think sometimes in form factor and I think sometimes in form factor and I think sometimes in doing everything for everyone mobile PCS doing everything for everyone mobile PCS doing everything for everyone mobile PCS Automotive someone um I see Tech Automotive someone um I see Tech Automotive someone um I see Tech journalists saying this I see you know journalists saying this I see you know journalists saying this I see you know the Twitter R will say oh it's just a the Twitter R will say oh it's just a the Twitter R will say oh it's just a phone chip on a PC or it's just a they phone chip on a PC or it's just a they phone chip on a PC or it's just a they kind of use a minimizing term it's just kind of use a minimizing term it's just kind of use a minimizing term it's just and I get that they're doing that to

and I get that they're doing that to and I get that they're doing that to understand what's happening like oh yeah understand what's happening like oh yeah understand what's happening like oh yeah they put the chip from the phone into they put the chip from the phone into they put the chip from the phone into the laptop how can I have that kind of the laptop how can I have that kind of the laptop how can I have that kind of power how can I have laptop power if power how can I have laptop power if power how can I have laptop power if it's quote just a phone chip and I know it's quote just a phone chip and I know it's quote just a phone chip and I know that's an that's an that's an oversimplification yeah and it's oversimplification yeah and it's oversimplification yeah and it's actually not true right so uh there are actually not true right so uh there are actually not true right so uh there are the chip that we build for the PC is the chip that we build for the PC is the chip that we build for the PC is very different from what we build in an very different from what we build in an very different from what we build in an auto is very different from what we put auto is very different from what we put auto is very different from what we put into a cell phone now there are lots of into a cell phone now there are lots of into a cell phone now there are lots of things that are shared like for example things that are shared like for example things that are shared like for example the CPU cord is shared right uh but even the CPU cord is shared right uh but even the CPU cord is shared right uh but even that there are various configurations of that there are various configurations of that there are various configurations of it right and then the numbers that go in it right and then the numbers that go in it right and then the numbers that go in the number of memory channels that the number of memory channels that the number of memory channels that you've got available the the number of you've got available the the number of you've got available the the number of accelerators and the like for example if accelerators and the like for example if accelerators and the like for example if you take the npu the npu internally has you take the npu the npu internally has you take the npu the npu internally has a number of processing elements pees in a number of processing elements pees in a number of processing elements pees in and yeah we expand those numbers and we and yeah we expand those numbers and we and yeah we expand those numbers and we reduce them based on the Target that we reduce them based on the Target that we reduce them based on the Target that we go after so they're not the same chips go after so they're not the same chips go after so they're not the same chips at all but there are the same at all but there are the same at all but there are the same building ah okay so then the Snapdragon building ah okay so then the Snapdragon building ah okay so then the Snapdragon X Series is not a phone chip that's X Series is not a phone chip that's X Series is not a phone chip that's something that is exclusively powering something that is exclusively powering something that is exclusively powering my PC I've got a a co-pilot plus PC my PC I've got a a co-pilot plus PC my PC I've got a a co-pilot plus PC right here that is a custom chip for right here that is a custom chip for right here that is a custom chip for that to make that PC a fantastic PC yes that to make that PC a fantastic PC yes that to make that PC a fantastic PC yes exactly right and it's very focused on exactly right and it's very focused on exactly right and it's very focused on that and it it's too big and it's too that and it it's too big and it's too that and it it's too big and it's too expensive to go into a cell phone right expensive to go into a cell phone right expensive to go into a cell phone right so in a cell phone you typically have so in a cell phone you typically have so in a cell phone you typically have something that is smaller uh and then it something that is smaller uh and then it something that is smaller uh and then it doesn't need those capabilities either

doesn't need those capabilities either doesn't need those capabilities either right so no they're very different ships right so no they're very different ships right so no they're very different ships but again same building blocks right so but again same building blocks right so but again same building blocks right so highly leveraged designs that's highly leveraged designs that's highly leveraged designs that's absolutely true uh and then on the absolutely true uh and then on the absolutely true uh and then on the software side too right we're very software side too right we're very software side too right we're very modular right so again there are modular right so again there are modular right so again there are different softare Stacks necessary for different softare Stacks necessary for different softare Stacks necessary for the PC there's a different software the PC there's a different software the PC there's a different software stack necessary for iot but again stack necessary for iot but again stack necessary for iot but again there's a lot of reuse right and we're there's a lot of reuse right and we're there's a lot of reuse right and we're trying to make as many things reusable trying to make as many things reusable trying to make as many things reusable because it reduces our development cost because it reduces our development cost because it reduces our development cost yeah now this is more of a question kind yeah now this is more of a question kind yeah now this is more of a question kind of about Vibes and I've always wanted to of about Vibes and I've always wanted to of about Vibes and I've always wanted to ask this because you know I'm on an ask this because you know I'm on an ask this because you know I'm on an Intel giant PC right now but when I Intel giant PC right now but when I Intel giant PC right now but when I throw my bag in the I throw my laptop in throw my bag in the I throw my laptop in throw my bag in the I throw my laptop in my bag I'm taking my arm my Qualcomm PC my bag I'm taking my arm my Qualcomm PC my bag I'm taking my arm my Qualcomm PC um it feels Snappy I've talked to people um it feels Snappy I've talked to people um it feels Snappy I've talked to people about this we don't understand why about this we don't understand why about this we don't understand why though there's a Vibe it's snappier you though there's a Vibe it's snappier you though there's a Vibe it's snappier you close it it shuts down you open it it close it it shuts down you open it it close it it shuts down you open it it turns on like all the things we were turns on like all the things we were turns on like all the things we were promised for PCs we actually are getting promised for PCs we actually are getting promised for PCs we actually are getting now like I don't put it in my bag and now like I don't put it in my bag and now like I don't put it in my bag and worry that it might become hot it um has worry that it might become hot it um has worry that it might become hot it um has this one I have only has 16 gigs of RAM this one I have only has 16 gigs of RAM this one I have only has 16 gigs of RAM which I thought would wouldn't be enough which I thought would wouldn't be enough which I thought would wouldn't be enough but it feels like a 32 gig ram machine but it feels like a 32 gig ram machine but it feels like a 32 gig ram machine running in running in running in 16 is there a technical explanation for 16 is there a technical explanation for 16 is there a technical explanation for why the Vibes are better on my arm-based why the Vibes are better on my arm-based why the Vibes are better on my arm-based PC well I think part of course is is is PC well I think part of course is is is PC well I think part of course is is is it's the design right uh that that we it's the design right uh that that we it's the design right uh that that we did I think but the reality is also but did I think but the reality is also but did I think but the reality is also but that applies both to the OS Windows OS that applies both to the OS Windows OS that applies both to the OS Windows OS uh as well as to uh the hardware a lot

uh as well as to uh the hardware a lot uh as well as to uh the hardware a lot of the Legacy stuff that you need to do of the Legacy stuff that you need to do of the Legacy stuff that you need to do on x 86 right uh you don't need to do on x 86 right uh you don't need to do on x 86 right uh you don't need to do right so you actually were able to look right so you actually were able to look right so you actually were able to look at these things in a different ways and at these things in a different ways and at these things in a different ways and and shed a lot of Legacy support that on and shed a lot of Legacy support that on and shed a lot of Legacy support that on x86 is is difficult to do by the way x86 is is difficult to do by the way x86 is is difficult to do by the way that's also the strength of x86 right if that's also the strength of x86 right if that's also the strength of x86 right if you've got a 40y old program and chances you've got a 40y old program and chances you've got a 40y old program and chances are it will still work but but arm you are it will still work but but arm you are it will still work but but arm you don't have that Legacy so you can start don't have that Legacy so you can start don't have that Legacy so you can start a fresh you don't need to support a fresh you don't need to support a fresh you don't need to support everything yeah that's a really great everything yeah that's a really great everything yeah that's a really great Point although I will say with uh just Point although I will say with uh just Point although I will say with uh just literally yesterday I was running a Dos literally yesterday I was running a Dos literally yesterday I was running a Dos box which is a Dos emulator which has box which is a Dos emulator which has box which is a Dos emulator which has been ported to arm and I was playing been ported to arm and I was playing been ported to arm and I was playing some dos games on it and and I was like some dos games on it and and I was like some dos games on it and and I was like the developers particularly the open the developers particularly the open the developers particularly the open source developers are kind of answering source developers are kind of answering source developers are kind of answering the call because the tool chain for the call because the tool chain for the call because the tool chain for building on arm is not as scary as building on arm is not as scary as building on arm is not as scary as people thought it would be a lot of people thought it would be a lot of people thought it would be a lot of folks are going oh well I remember the folks are going oh well I remember the folks are going oh well I remember the shift from x86 to shift from x86 to shift from x86 to x64 and some of that recompilation and x64 and some of that recompilation and x64 and some of that recompilation and they're like ah I don't want to they're like ah I don't want to they're like ah I don't want to recompile for arm and they're like oh it recompile for arm and they're like oh it recompile for arm and they're like oh it just worked are you going after like just worked are you going after like just worked are you going after like open- Source folks and saying look look open- Source folks and saying look look open- Source folks and saying look look how easy it is to to rec arm so we're how easy it is to to rec arm so we're how easy it is to to rec arm so we're doing that and we're also porting a lot doing that and we're also porting a lot doing that and we're also porting a lot of the open source tools right so llbm of the open source tools right so llbm of the open source tools right so llbm and things like that we've been actively and things like that we've been actively and things like that we've been actively working on to get a Windows version working on to get a Windows version working on to get a Windows version Windows obviously is a little bit Windows obviously is a little bit Windows obviously is a little bit different than than Linux abis are a different than than Linux abis are a different than than Linux abis are a little bit different so we we build we little bit different so we we build we little bit different so we we build we work together with Microsoft on getting

work together with Microsoft on getting work together with Microsoft on getting L and little L and little L and little was Microsoft Laro and and in qualcom we was Microsoft Laro and and in qualcom we was Microsoft Laro and and in qualcom we actually worked on getting an llm actually worked on getting an llm actually worked on getting an llm version for Windows stud and then make version for Windows stud and then make version for Windows stud and then make sure that all the optimizations are in sure that all the optimizations are in sure that all the optimizations are in there obviously so yeah we're actively there obviously so yeah we're actively there obviously so yeah we're actively going after that of course another big going after that of course another big going after that of course another big thing that helps on the arm side is the thing that helps on the arm side is the thing that helps on the arm side is the arm 64 EC encoding right which was done arm 64 EC encoding right which was done arm 64 EC encoding right which was done again by the windows team and that's to again by the windows team and that's to again by the windows team and that's to make it such that you can have armed make it such that you can have armed make it such that you can have armed binaries but then still call into x86 binaries but then still call into x86 binaries but then still call into x86 binaries which are then emulated and binaries which are then emulated and binaries which are then emulated and this has all to do with the transition this has all to do with the transition this has all to do with the transition so the transition uh there's been a lot so the transition uh there's been a lot so the transition uh there's been a lot of focus primarily by by Microsoft to to of focus primarily by by Microsoft to to of focus primarily by by Microsoft to to figure out how can we make this figure out how can we make this figure out how can we make this transition as as as painless as possible transition as as as painless as possible transition as as as painless as possible right because changing isas is difficult right because changing isas is difficult right because changing isas is difficult right it takes time you're calling out right it takes time you're calling out right it takes time you're calling out the 64-bit transition uh and yeah it the 64-bit transition uh and yeah it the 64-bit transition uh and yeah it took about 10 years before the industry took about 10 years before the industry took about 10 years before the industry had really adopted that yeah and again had really adopted that yeah and again had really adopted that yeah and again again in in my uh in our podcast again in in my uh in our podcast again in in my uh in our podcast tradition of expanding uh acronyms EC tradition of expanding uh acronyms EC tradition of expanding uh acronyms EC means emulation compatible so arm64 EC means emulation compatible so arm64 EC means emulation compatible so arm64 EC it's an ABI an application binary it's an ABI an application binary it's an ABI an application binary interface it's basically allowing an interface it's basically allowing an interface it's basically allowing an arm64 binary to run natively but you can arm64 binary to run natively but you can arm64 binary to run natively but you can interupt with x64 which is so convenient interupt with x64 which is so convenient interupt with x64 which is so convenient because you might have an entire because you might have an entire because you might have an entire application with a little chunk of x64 application with a little chunk of x64 application with a little chunk of x64 and you need to figure out some calling and you need to figure out some calling and you need to figure out some calling conventions to make that work that would conventions to make that work that would conventions to make that work that would be a hassle but it just isn't anymore

be a hassle but it just isn't anymore be a hassle but it just isn't anymore because of that that API well for because of that that API well for because of that that API well for example if you've got uh a a video example if you've got uh a a video example if you've got uh a a video player and you need an x86 you need a player and you need an x86 you need a player and you need an x86 you need a codec that's only available on x86 right codec that's only available on x86 right codec that's only available on x86 right a dll you can basically plug it in uh if a dll you can basically plug it in uh if a dll you can basically plug it in uh if you adhere to that particular ABI you adhere to that particular ABI you adhere to that particular ABI so um I would have to say maybe 5 10 so um I would have to say maybe 5 10 so um I would have to say maybe 5 10 years ago like I knew of Qualcomm but I years ago like I knew of Qualcomm but I years ago like I knew of Qualcomm but I wouldn't bump into Qualcomm on the wouldn't bump into Qualcomm on the wouldn't bump into Qualcomm on the street but now you see Qualcomm street but now you see Qualcomm street but now you see Qualcomm Engineers uh you know in community Engineers uh you know in community Engineers uh you know in community channels in slacks and discords on stack channels in slacks and discords on stack channels in slacks and discords on stack Overflow uh starting to show up in Overflow uh starting to show up in Overflow uh starting to show up in GitHub issues as well which I think is GitHub issues as well which I think is GitHub issues as well which I think is really cool like Microsoft has done that really cool like Microsoft has done that really cool like Microsoft has done that for a while you don't usually see apple for a while you don't usually see apple for a while you don't usually see apple Engineers kind of like wandering around Engineers kind of like wandering around Engineers kind of like wandering around talking to people was that a a shift talking to people was that a a shift talking to people was that a a shift like a mental shift for Qualcomm like a mental shift for Qualcomm like a mental shift for Qualcomm Engineers to start showing up in the Engineers to start showing up in the Engineers to start showing up in the community absolutely um it was uh both a community absolutely um it was uh both a community absolutely um it was uh both a shift in thinking right uh and and it shift in thinking right uh and and it shift in thinking right uh and and it wasn't just on the engineering site um wasn't just on the engineering site um wasn't just on the engineering site um it was also on the legal site right and it was also on the legal site right and it was also on the legal site right and on our licensing site people needed to on our licensing site people needed to on our licensing site people needed to get comfortable with like Hey we're get comfortable with like Hey we're get comfortable with like Hey we're talking about our technology uh because talking about our technology uh because talking about our technology uh because as I said as an IP company you're very as I said as an IP company you're very as I said as an IP company you're very conservative the easiest thing is to be conservative the easiest thing is to be conservative the easiest thing is to be very conservative your um and that works very conservative your um and that works very conservative your um and that works well right if your ecosystem is well right if your ecosystem is well right if your ecosystem is relatively small and closed but relatively small and closed but relatively small and closed but obviously we're now working in all these obviously we're now working in all these obviously we're now working in all these open ecosystems and ecosystems that are open ecosystems and ecosystems that are open ecosystems and ecosystems that are established right they have their their

established right they have their their established right they have their their rules of engagements and so we need to rules of engagements and so we need to rules of engagements and so we need to conform to that as opposed to us trying conform to that as opposed to us trying conform to that as opposed to us trying to say hey this is how it work in the to say hey this is how it work in the to say hey this is how it work in the cell phone space do it like this because cell phone space do it like this because cell phone space do it like this because nobody cares right people want to say nobody cares right people want to say nobody cares right people want to say well we're already doing it this way well we're already doing it this way well we're already doing it this way this works for us this works with all this works for us this works with all this works for us this works with all these other vendors welcome you go and these other vendors welcome you go and these other vendors welcome you go and do the same thing so yeah so we've been do the same thing so yeah so we've been do the same thing so yeah so we've been very much engaging with the community to very much engaging with the community to very much engaging with the community to your point Discord uh GitHub uh taken a your point Discord uh GitHub uh taken a your point Discord uh GitHub uh taken a much more public stance much more public stance much more public stance toward Linux like for example although toward Linux like for example although toward Linux like for example although it wasn't part of the product plan right it wasn't part of the product plan right it wasn't part of the product plan right the engineers did actually enable Linux the engineers did actually enable Linux the engineers did actually enable Linux on on uh on the the ex Elite but we on on uh on the the ex Elite but we on on uh on the the ex Elite but we basically and we released all of that we basically and we released all of that we basically and we released all of that we open source all of that so people are open source all of that so people are open source all of that so people are now building Linux for the various uh I now building Linux for the various uh I now building Linux for the various uh I think D XPS Lenovo platforms um and yeah think D XPS Lenovo platforms um and yeah think D XPS Lenovo platforms um and yeah you can just download deut that's so you can just download deut that's so you can just download deut that's so cool I really think the part that's cool I really think the part that's cool I really think the part that's exciting about it is like you said the exciting about it is like you said the exciting about it is like you said the building blocks that like well yeah why building blocks that like well yeah why building blocks that like well yeah why shouldn't they be able to play with shouldn't they be able to play with shouldn't they be able to play with these building blocks like I bought the these building blocks like I bought the these building blocks like I bought the Silicon I want to I have this machine I Silicon I want to I have this machine I Silicon I want to I have this machine I want to see what I can do with it uh the want to see what I can do with it uh the want to see what I can do with it uh the other thing that I think is interesting other thing that I think is interesting other thing that I think is interesting is it's not just when we say system on a is it's not just when we say system on a is it's not just when we say system on a chip for some people it's like it just chip for some people it's like it just chip for some people it's like it just works like for my non-technical family works like for my non-technical family works like for my non-technical family members they're going to pick up one of members they're going to pick up one of members they're going to pick up one of these uh you know Snapdragon based PCS these uh you know Snapdragon based PCS these uh you know Snapdragon based PCS and it's going to just work and they're and it's going to just work and they're and it's going to just work and they're going to love it and they're not going going to love it and they're not going going to love it and they're not going to even know like my non-technical wife to even know like my non-technical wife to even know like my non-technical wife is not going to care as long as her apps

is not going to care as long as her apps is not going to care as long as her apps run but when I look at it I go into task run but when I look at it I go into task run but when I look at it I go into task manager and I'm like wow there's a manager and I'm like wow there's a manager and I'm like wow there's a system on a chip but we've got an npu A system on a chip but we've got an npu A system on a chip but we've got an npu A GPU and a CPU it's a it's a GPU and a CPU it's a it's a GPU and a CPU it's a it's a heterogeneous compute heterogeneous compute heterogeneous compute model on a chip which I think is so model on a chip which I think is so model on a chip which I think is so amazing yeah but to be honest all all amazing yeah but to be honest all all amazing yeah but to be honest all all Every semiconductor is going to an S so Every semiconductor is going to an S so Every semiconductor is going to an S so right when when I was at AMD we were right when when I was at AMD we were right when when I was at AMD we were building at s so we wouldn't call it building at s so we wouldn't call it building at s so we wouldn't call it that back then right really by the time that back then right really by the time that back then right really by the time when we started to integrate remember when we started to integrate remember when we started to integrate remember North Bridges and South Bridges right North Bridges and South Bridges right North Bridges and South Bridges right where they they became one that was an S where they they became one that was an S where they they became one that was an S so right and so in that sense everything so right and so in that sense everything so right and so in that sense everything really is an S so nowadays but I think really is an S so nowadays but I think really is an S so nowadays but I think your most important thing with comment your most important thing with comment your most important thing with comment about excelite is it just works right about excelite is it just works right about excelite is it just works right Windows comes up just like any other Windows comes up just like any other Windows comes up just like any other system it works in the same way you've system it works in the same way you've system it works in the same way you've got the same tools running got the same tools running got the same tools running uh I think the Microsoft folks did a uh I think the Microsoft folks did a uh I think the Microsoft folks did a phenomenal job at the binary emulation phenomenal job at the binary emulation phenomenal job at the binary emulation and also improved the performance of and also improved the performance of and also improved the performance of that so again a lot of those programs that so again a lot of those programs that so again a lot of those programs will just will just will just work yeah it's funny because as someone work yeah it's funny because as someone work yeah it's funny because as someone who myself who keeps task manager open who myself who keeps task manager open who myself who keeps task manager open on another monitor at all times you know on another monitor at all times you know on another monitor at all times you know like I think there's a certain like I think there's a certain like I think there's a certain personality type that is always running personality type that is always running personality type that is always running task manager when I started with the task manager when I started with the task manager when I started with the early early arm developer kits I would early early arm developer kits I would early early arm developer kits I would have task manager open and I would have have task manager open and I would have have task manager open and I would have that one column that shows me whether or that one column that shows me whether or that one column that shows me whether or not this process is an x64 process or an not this process is an x64 process or an not this process is an x64 process or an arm process so I would have this

arm process so I would have this arm process so I would have this instruction set anxiety of a kind and instruction set anxiety of a kind and instruction set anxiety of a kind and that has slowly Fallen away because of that has slowly Fallen away because of that has slowly Fallen away because of tools on the Windows side like like wind tools on the Windows side like like wind tools on the Windows side like like wind get which is like apt get but for get which is like apt get but for get which is like apt get but for Windows when I say wind get install it's Windows when I say wind get install it's Windows when I say wind get install it's going to get me the right version the going to get me the right version the going to get me the right version the arm version when I went recently and got arm version when I went recently and got arm version when I went recently and got LM Studio which is one of my most LM Studio which is one of my most LM Studio which is one of my most favorite AI apps there 's a Snapdragon favorite AI apps there 's a Snapdragon favorite AI apps there 's a Snapdragon version of LM studio and it I don't version of LM studio and it I don't version of LM studio and it I don't think about the instruction set anymore think about the instruction set anymore think about the instruction set anymore everything's just working which is super everything's just working which is super everything's just working which is super cool and and that's how it should be cool and and that's how it should be cool and and that's how it should be right I think that the ISA is really not right I think that the ISA is really not right I think that the ISA is really not that relevant right it it's what you run that relevant right it it's what you run that relevant right it it's what you run on top of it and it's what you build on on top of it and it's what you build on on top of it and it's what you build on top of it uh and then yeah the underly top of it uh and then yeah the underly top of it uh and then yeah the underly processor needs to stand on its own processor needs to stand on its own processor needs to stand on its own Merit which really is about performance Merit which really is about performance Merit which really is about performance it's about bettery Lifetime right in in it's about bettery Lifetime right in in it's about bettery Lifetime right in in a notebook format it's about what is a notebook format it's about what is a notebook format it's about what is your your video performance what is your your video performance what is your your video performance what is about your your AI performance right about your your AI performance right about your your AI performance right those are the kind of things you should those are the kind of things you should those are the kind of things you should measure something on yeah absolutely so measure something on yeah absolutely so measure something on yeah absolutely so um when we're thinking about doing AI um um when we're thinking about doing AI um um when we're thinking about doing AI um I have talked to uh to again I have talked to uh to again I have talked to uh to again non-technical people who think that AI non-technical people who think that AI non-technical people who think that AI is just sending your data off to some is just sending your data off to some is just sending your data off to some giant company and some very large giant company and some very large giant company and some very large language model doing the work but I language model doing the work but I language model doing the work but I think more and more and more we're going think more and more and more we're going think more and more and more we're going to have expectations for local private to have expectations for local private to have expectations for local private processing and AI processing happening processing and AI processing happening processing and AI processing happening on an airplane I always think about

on an airplane I always think about on an airplane I always think about airplane mode like what AI scenarios can airplane mode like what AI scenarios can airplane mode like what AI scenarios can I do while I'm on the plane and that npu I do while I'm on the plane and that npu I do while I'm on the plane and that npu is where that that comes in uh into play is where that that comes in uh into play is where that that comes in uh into play Lally agree I think you're calling out Lally agree I think you're calling out Lally agree I think you're calling out privacy which is obviously one uh privacy which is obviously one uh privacy which is obviously one uh response time is another reason why you response time is another reason why you response time is another reason why you want to do this uh uh locally right want to do this uh uh locally right want to do this uh uh locally right because that's where uh you don't have because that's where uh you don't have because that's where uh you don't have to go into the cloud you don't have to to go into the cloud you don't have to to go into the cloud you don't have to wait for milliseconds to go to Something wait for milliseconds to go to Something wait for milliseconds to go to Something in the cloud and you answer back it in the cloud and you answer back it in the cloud and you answer back it responds immediately and then of course responds immediately and then of course responds immediately and then of course there is power efficiency I think the there is power efficiency I think the there is power efficiency I think the the the cloud providers are finding out the the cloud providers are finding out the the cloud providers are finding out that buying WhatsApp gpus right um is that buying WhatsApp gpus right um is that buying WhatsApp gpus right um is expensive it's very power hungry and so expensive it's very power hungry and so expensive it's very power hungry and so offloading it to the devices themselves offloading it to the devices themselves offloading it to the devices themselves especi especially when they've got especi especially when they've got especi especially when they've got efficient accelerators right is a much efficient accelerators right is a much efficient accelerators right is a much better tradeoff with as well now better tradeoff with as well now better tradeoff with as well now obviously you can't run the same like obviously you can't run the same like obviously you can't run the same like half a billion parameter models there half a billion parameter models there half a billion parameter models there right so you're talking about 7 billion right so you're talking about 7 billion right so you're talking about 7 billion parameter parameter parameter well she I might metri WR but you can well she I might metri WR but you can well she I might metri WR but you can run 7 billion paramet models right on run 7 billion paramet models right on run 7 billion paramet models right on the in vi just fine uh but yeah if you the in vi just fine uh but yeah if you the in vi just fine uh but yeah if you want to run an 80 billion dollar 80 want to run an 80 billion dollar 80 want to run an 80 billion dollar 80 billion parameter model then it becomes billion parameter model then it becomes billion parameter model then it becomes more tricky right so I think there more tricky right so I think there more tricky right so I think there you're going to see these Hybrid models you're going to see these Hybrid models you're going to see these Hybrid models where uh there is certainly a certain where uh there is certainly a certain where uh there is certainly a certain set of that that all the queries most of set of that that all the queries most of set of that that all the queries most of the queries are first processed locally

the queries are first processed locally the queries are first processed locally and then they are pushed into the cloud and then they are pushed into the cloud and then they are pushed into the cloud if the local model is not capable of if the local model is not capable of if the local model is not capable of answering it or dealing with it right so answering it or dealing with it right so answering it or dealing with it right so you're going to see this this kind of you're going to see this this kind of you're going to see this this kind of trade off and of course in the meantime trade off and of course in the meantime trade off and of course in the meantime I also fully expect that these I also fully expect that these I also fully expect that these accelerators will continue to grow and accelerators will continue to grow and accelerators will continue to grow and increase in capabilities and also the increase in capabilities and also the increase in capabilities and also the models will change right we've had some models will change right we've had some models will change right we've had some very interesting disruptive changes just very interesting disruptive changes just very interesting disruptive changes just over the last couple years where the over the last couple years where the over the last couple years where the models completely changed so is the models completely changed so is the models completely changed so is the future really going to be one large future really going to be one large future really going to be one large language model or is it going to be lots language model or is it going to be lots language model or is it going to be lots of little models with an orchestrator on of little models with an orchestrator on of little models with an orchestrator on top of it and that again is going to top of it and that again is going to top of it and that again is going to change this architecture completely and change this architecture completely and change this architecture completely and what you can do on a client device what you can do on a client device what you can do on a client device changes completely yeah yeah I think changes completely yeah yeah I think changes completely yeah yeah I think that that that kind of realization that that that that kind of realization that that that that kind of realization that these very large language models were these very large language models were these very large language models were really kind of a melding of the minds of really kind of a melding of the minds of really kind of a melding of the minds of many little models and that you can get many little models and that you can get many little models and that you can get really really effective results with a really really effective results with a really really effective results with a couple of single-digit billion parameter couple of single-digit billion parameter couple of single-digit billion parameter models rather than one super or Uber models rather than one super or Uber models rather than one super or Uber model and clearly Microsoft and model and clearly Microsoft and model and clearly Microsoft and Microsoft research is at the Forefront Microsoft research is at the Forefront Microsoft research is at the Forefront of this right they're doing a lot of of this right they're doing a lot of of this right they're doing a lot of work here and that's actually a very work here and that's actually a very work here and that's actually a very interesting development from my interesting development from my interesting development from my perspective because then all of a sudden perspective because then all of a sudden perspective because then all of a sudden you can run these mod you can get the you can run these mod you can get the you can run these mod you can get the experience of a large language model experience of a large language model experience of a large language model really large model on energy really large model on energy really large model on energy constrainted devices right like a BC constrainted devices right like a BC constrainted devices right like a BC like a cell phone and again goes back to like a cell phone and again goes back to like a cell phone and again goes back to what I think is the most interesting what I think is the most interesting what I think is the most interesting thing about AI it's going to be the way thing about AI it's going to be the way thing about AI it's going to be the way it's going to fundamentally change the it's going to fundamentally change the it's going to fundamentally change the way how we interact with devices it's

way how we interact with devices it's way how we interact with devices it's going to be a much more natural going to be a much more natural going to be a much more natural experience you're going to be talking to experience you're going to be talking to experience you're going to be talking to it I mean where we're currently being it I mean where we're currently being it I mean where we're currently being handicapped by keyboard right you now handicapped by keyboard right you now handicapped by keyboard right you now you're going to be able to talk to the you're going to be able to talk to the you're going to be able to talk to the device and you can interact with it I uh device and you can interact with it I uh device and you can interact with it I uh I I I remember in I I I remember in I I I remember in 1984 my parents took me out of school 1984 my parents took me out of school 1984 my parents took me out of school and we went to see Star Trek 4 the when and we went to see Star Trek 4 the when and we went to see Star Trek 4 the when it was when it was launched and that's it was when it was launched and that's it was when it was launched and that's the one where they go back in time to the one where they go back in time to the one where they go back in time to save the whales and there's a scene save the whales and there's a scene save the whales and there's a scene where Montgomery Scott the Engineer sits where Montgomery Scott the Engineer sits where Montgomery Scott the Engineer sits down and they ask him to go and do some down and they ask him to go and do some down and they ask him to go and do some Computing and he sits down on an Apple Computing and he sits down on an Apple Computing and he sits down on an Apple computer and he picks up the mouse and computer and he picks up the mouse and computer and he picks up the mouse and he goes computer and they say use the he goes computer and they say use the he goes computer and they say use the keyboard and he's like oh how quaint and keyboard and he's like oh how quaint and keyboard and he's like oh how quaint and I finally feel like we're there where I I finally feel like we're there where I I finally feel like we're there where I could you know sit down to a computer could you know sit down to a computer could you know sit down to a computer and maybe have a whole ongoing and maybe have a whole ongoing and maybe have a whole ongoing conversation it is going to change the conversation it is going to change the conversation it is going to change the interaction model fundamentally and it's interaction model fundamentally and it's interaction model fundamentally and it's only taken you know 40 years yeah oh only taken you know 40 years yeah oh only taken you know 40 years yeah oh it's going really fast yeah now that it's going really fast yeah now that it's going really fast yeah now that said I think that AI still needs to come said I think that AI still needs to come said I think that AI still needs to come a long way right I mean hallucinations a long way right I mean hallucinations a long way right I mean hallucinations are real and so AI is a great tool in my are real and so AI is a great tool in my are real and so AI is a great tool in my mind but you need to be an expert uh so mind but you need to be an expert uh so mind but you need to be an expert uh so that you can spot those hallucinations that you can spot those hallucinations that you can spot those hallucinations and actually correct that the um one and actually correct that the um one and actually correct that the um one other thing I wanted to kind of dig into other thing I wanted to kind of dig into other thing I wanted to kind of dig into a little bit is the the kinds of a little bit is the the kinds of a little bit is the the kinds of instruction sets that a CPU is good at instruction sets that a CPU is good at instruction sets that a CPU is good at versus a GPU versus an npu because a lot versus a GPU versus an npu because a lot versus a GPU versus an npu because a lot of us went out and bought big expensive of us went out and bought big expensive of us went out and bought big expensive gpus thinking that that was going to be gpus thinking that that was going to be gpus thinking that that was going to be our way forward to run local models to

our way forward to run local models to our way forward to run local models to do that kind of work and then npus are do that kind of work and then npus are do that kind of work and then npus are starting to come up and I don't think starting to come up and I don't think starting to come up and I don't think about them in the same way as a big about them in the same way as a big about them in the same way as a big heavy heating the room kind of video heavy heating the room kind of video heavy heating the room kind of video card what are gpus good at from a math card what are gpus good at from a math card what are gpus good at from a math perspective and what are npus good at perspective and what are npus good at perspective and what are npus good at and why is one better at some parts of and why is one better at some parts of and why is one better at some parts of AI versus other parts of AI so the the AI versus other parts of AI so the the AI versus other parts of AI so the the biggest thing is the gpus do everything biggest thing is the gpus do everything biggest thing is the gpus do everything in floating point right and floating in floating point right and floating in floating point right and floating points are actually complicated points are actually complicated points are actually complicated operation require lots of transistors to operation require lots of transistors to operation require lots of transistors to actually make them happen but one of the actually make them happen but one of the actually make them happen but one of the big things you see with the npu is we do big things you see with the npu is we do big things you see with the npu is we do everything in integers and integers are everything in integers and integers are everything in integers and integers are much simpler require much fewer much simpler require much fewer much simpler require much fewer transistors and therefore less power transistors and therefore less power transistors and therefore less power significant power but there is a significant power but there is a significant power but there is a conversion step because most models are conversion step because most models are conversion step because most models are trained all models are really trained trained all models are really trained trained all models are really trained with floting coins and then you have to with floting coins and then you have to with floting coins and then you have to retrain them or what what re quantize I retrain them or what what re quantize I retrain them or what what re quantize I should use the correct terminology into should use the correct terminology into should use the correct terminology into integers and that is a tricky operation integers and that is a tricky operation integers and that is a tricky operation and that is time consuming but once you and that is time consuming but once you and that is time consuming but once you do that you finish that then you've got do that you finish that then you've got do that you finish that then you've got an amazing performance uh and an amazing an amazing performance uh and an amazing an amazing performance uh and an amazing power profile out of running those power profile out of running those power profile out of running those models so that is really and there are models so that is really and there are models so that is really and there are some other architectural differences too some other architectural differences too some other architectural differences too but those are the big things but those are the big things but those are the big things you also refer to a generic uh did you you also refer to a generic uh did you you also refer to a generic uh did you say generic CPU or you talk just kind of say generic CPU or you talk just kind of say generic CPU or you talk just kind of CPUs in a generic sense uh versus you CPUs in a generic sense uh versus you CPUs in a generic sense uh versus you know like now I have three processing

know like now I have three processing know like now I have three processing units you know in my in my machine I units you know in my in my machine I units you know in my in my machine I have three different uh tools available have three different uh tools available have three different uh tools available to me in my toolkit in one in one system to me in my toolkit in one in one system to me in my toolkit in one in one system on a chip and to be honest they will all on a chip and to be honest they will all on a chip and to be honest they will all be able to run models it's just a matter be able to run models it's just a matter be able to run models it's just a matter of how efficient right so and now there of how efficient right so and now there of how efficient right so and now there is an interesting thing to it right if is an interesting thing to it right if is an interesting thing to it right if you look at your npu it is more you look at your npu it is more you look at your npu it is more efficient than your GPU your GPU is more efficient than your GPU your GPU is more efficient than your GPU your GPU is more efficient than your CPU now they can all efficient than your CPU now they can all efficient than your CPU now they can all run these models but that's an run these models but that's an run these models but that's an efficiency thing now from a developer efficiency thing now from a developer efficiency thing now from a developer perspective it's actually the other way perspective it's actually the other way perspective it's actually the other way around it's much easier to bring up a around it's much easier to bring up a around it's much easier to bring up a model on a CPU that it is on a GPU that model on a CPU that it is on a GPU that model on a CPU that it is on a GPU that it is on an npu right so it's all about it is on an npu right so it's all about it is on an npu right so it's all about like hey are you willing to take your like hey are you willing to take your like hey are you willing to take your time Investments to get something time Investments to get something time Investments to get something running on the npu now the returns of running on the npu now the returns of running on the npu now the returns of that is she have an amazing power in and that is she have an amazing power in and that is she have an amazing power in and performance profile uh but it's going to performance profile uh but it's going to performance profile uh but it's going to take you time to get it there and so if take you time to get it there and so if take you time to get it there and so if time to Market is really critical you'll time to Market is really critical you'll time to Market is really critical you'll see people go first for the CPU then to see people go first for the CPU then to see people go first for the CPU then to the GPU because there's a lot of Open the GPU because there's a lot of Open the GPU because there's a lot of Open Source tools available for gpus and then Source tools available for gpus and then Source tools available for gpus and then you see them going to the npu so I I you see them going to the npu so I I you see them going to the npu so I I think all three are going to play a role think all three are going to play a role think all three are going to play a role uh it all depends on really what your uh it all depends on really what your uh it all depends on really what your parameters are time to Market is parameters are time to Market is parameters are time to Market is critical to you you'll probably focus on critical to you you'll probably focus on critical to you you'll probably focus on the CPU right if low power is critical the CPU right if low power is critical the CPU right if low power is critical you're going to focus on you're going to focus on you're going to focus on the that makes a lot of sense so then the that makes a lot of sense so then the that makes a lot of sense so then for my you know popping back off the

for my you know popping back off the for my you know popping back off the stack to my conversation around LM stack to my conversation around LM stack to my conversation around LM Studio which is a great application that Studio which is a great application that Studio which is a great application that we can use on any PC I'm using it we can use on any PC I'm using it we can use on any PC I'm using it currently on my Snapdragon X Elite PC currently on my Snapdragon X Elite PC currently on my Snapdragon X Elite PC the intent will be for smaller parameter the intent will be for smaller parameter the intent will be for smaller parameter language models you maybe seven or eight language models you maybe seven or eight language models you maybe seven or eight billion rather than 80 to run on that on billion rather than 80 to run on that on billion rather than 80 to run on that on that low power npu which means I'm going that low power npu which means I'm going that low power npu which means I'm going to get great battery life as opposed to to get great battery life as opposed to to get great battery life as opposed to firing up and heating up a GPU to do firing up and heating up a GPU to do firing up and heating up a GPU to do similar work yeah like take for example similar work yeah like take for example similar work yeah like take for example things like another type of generative things like another type of generative things like another type of generative AI model is Tiny whisper right which AI model is Tiny whisper right which AI model is Tiny whisper right which basically does speech to a text basically does speech to a text basically does speech to a text conversion it's about one one and a half conversion it's about one one and a half conversion it's about one one and a half billion or salate that's a perfect billion or salate that's a perfect billion or salate that's a perfect candidate to run on an npu right uh candidate to run on an npu right uh candidate to run on an npu right uh because it needs need to be on all the because it needs need to be on all the because it needs need to be on all the time it needs to listen all the time you time it needs to listen all the time you time it needs to listen all the time you don't want to spend the majority of your don't want to spend the majority of your don't want to spend the majority of your battery Lifetime on it right you want to battery Lifetime on it right you want to battery Lifetime on it right you want to take it as little power as possible so take it as little power as possible so take it as little power as possible so yeah great application of uh of the npu yeah great application of uh of the npu yeah great application of uh of the npu but we're also running w 7 billion but we're also running w 7 billion but we're also running w 7 billion parameter models on it right so there's parameter models on it right so there's parameter models on it right so there's there's a wide scale of things that you there's a wide scale of things that you there's a wide scale of things that you can those npus are very capable much can those npus are very capable much can those npus are very capable much more capable than a lot of people assume more capable than a lot of people assume more capable than a lot of people assume I remember the first conversations we I remember the first conversations we I remember the first conversations we had with customers about like yeah we had with customers about like yeah we had with customers about like yeah we got Lama 7 billion up running on the npu got Lama 7 billion up running on the npu got Lama 7 billion up running on the npu where they were like what like we're where they were like what like we're where they were like what like we're running that on on our GPU clusters it's running that on on our GPU clusters it's running that on on our GPU clusters it's like you got it running on your little like you got it running on your little like you got it running on your little uh s so in a cell phone and answer is uh s so in a cell phone and answer is uh s so in a cell phone and answer is yes we can't you know that is a really

yes we can't you know that is a really yes we can't you know that is a really great point I would have to say I am great point I would have to say I am great point I would have to say I am also guilty of that where I presumed also guilty of that where I presumed also guilty of that where I presumed maybe a year or two ago that the npu maybe a year or two ago that the npu maybe a year or two ago that the npu would be a toy but then I saw uh this would be a toy but then I saw uh this would be a toy but then I saw uh this company called seph which is an company called seph which is an company called seph which is an accessibility company running a model on accessibility company running a model on accessibility company running a model on the npu full-time for accessibility such the npu full-time for accessibility such the npu full-time for accessibility such that someone who is disabled or that someone who is disabled or that someone who is disabled or differently abled could interact with differently abled could interact with differently abled could interact with their PC and not take up CPU time like their PC and not take up CPU time like their PC and not take up CPU time like why should they pay a tax of 10 20 30% why should they pay a tax of 10 20 30% why should they pay a tax of 10 20 30% of their CPU to you know have a model of their CPU to you know have a model of their CPU to you know have a model that's running entirely locally looking that's running entirely locally looking that's running entirely locally looking at their face allowing to them to at their face allowing to them to at their face allowing to them to interact with their with their PC that's interact with their with their PC that's interact with their with their PC that's a perfect example of a co-processor and a perfect example of a co-processor and a perfect example of a co-processor and it's also a good example of how the user it's also a good example of how the user it's also a good example of how the user interface with our computer is going to interface with our computer is going to interface with our computer is going to change right yeah it's a great Point change right yeah it's a great Point change right yeah it's a great Point whether it be speech or your face or whether it be speech or your face or whether it be speech or your face or whatever you know everyone should be whatever you know everyone should be whatever you know everyone should be able to have those experiences without able to have those experiences without able to have those experiences without using up all their battery life and all using up all their battery life and all using up all their battery life and all their power fantastic okay cool well I their power fantastic okay cool well I their power fantastic okay cool well I so you can get a Snapdragon Elite device so you can get a Snapdragon Elite device so you can get a Snapdragon Elite device like in store like this is not a theory like in store like this is not a theory like in store like this is not a theory you can go down to Costco you can go to you can go down to Costco you can go to you can go down to Costco you can go to Amazon you can go to Best Buy and buy Amazon you can go to Best Buy and buy Amazon you can go to Best Buy and buy these things right now they're on the these things right now they're on the these things right now they're on the Shelf that's so Cool's number of oebs so Shelf that's so Cool's number of oebs so Shelf that's so Cool's number of oebs so that's a cool thing too it's not just that's a cool thing too it's not just that's a cool thing too it's not just Dell HB Dell HB Dell HB Lenovo you can get they they we launched Lenovo you can get they they we launched Lenovo you can get they they we launched with a lot of Partners here so yeah it's with a lot of Partners here so yeah it's with a lot of Partners here so yeah it's it's broadly available as well as you it's broadly available as well as you it's broadly available as well as you said your lenaro collaboration so said your lenaro collaboration so said your lenaro collaboration so there's documentation and tools for

there's documentation and tools for there's documentation and tools for developers to run Linux on these Snap developers to run Linux on these Snap developers to run Linux on these Snap Dragon x uh devices as well we've been Dragon x uh devices as well we've been Dragon x uh devices as well we've been working with Laro both on Windows tools working with Laro both on Windows tools working with Laro both on Windows tools that run natively on Windows but also of that run natively on Windows but also of that run natively on Windows but also of course Linux right that's a big part of course Linux right that's a big part of course Linux right that's a big part of of of what lenaro does and so yeah of of what lenaro does and so yeah of of what lenaro does and so yeah that's up and running on these systems that's up and running on these systems that's up and running on these systems too absolutely fantastic well thank you too absolutely fantastic well thank you too absolutely fantastic well thank you so much uh Lander vanor senior vice so much uh Lander vanor senior vice so much uh Lander vanor senior vice president at Qualcomm for chatting with president at Qualcomm for chatting with president at Qualcomm for chatting with me today you're welcome hey I wanted to me today you're welcome hey I wanted to me today you're welcome hey I wanted to uh remind people of one more thing we uh remind people of one more thing we uh remind people of one more thing we have a joint uh webinar with Microsoft have a joint uh webinar with Microsoft have a joint uh webinar with Microsoft coming up at the 17th right uh so really coming up at the 17th right uh so really coming up at the 17th right uh so really want to encourage people to go there and want to encourage people to go there and want to encourage people to go there and and and register for that at welcome.com and and register for that at welcome.com and and register for that at welcome.com and developer uh so yeah um and in and developer uh so yeah um and in and developer uh so yeah um and in general I think going to that website general I think going to that website general I think going to that website you'll find a lot of developer tools and you'll find a lot of developer tools and you'll find a lot of developer tools and a lot more information out platforms not a lot more information out platforms not a lot more information out platforms not just the PC uh ex Elite but also around just the PC uh ex Elite but also around just the PC uh ex Elite but also around iot and all the other things we kind of iot and all the other things we kind of iot and all the other things we kind of De so thank you very much for this De so thank you very much for this De so thank you very much for this conversation all right folks can check conversation all right folks can check conversation all right folks can check out out out qualcomm.com developer for all things qualcomm.com developer for all things qualcomm.com developer for all things Qualcomm all things developer as well as Qualcomm all things developer as well as Qualcomm all things developer as well as registering for that webinar this has registering for that webinar this has registering for that webinar this has been another episode of Hansel minutes been another episode of Hansel minutes been another episode of Hansel minutes and we'll see you again next week a a [Music]