I Put the M5 MacBook Air Against Every Generation for Dev Work

This is five generations of Apple This is five generations of Apple Silicon MacBook Air. And every year, Silicon MacBook Air. And every year, Silicon MacBook Air. And every year, Apple made it faster and faster while Apple made it faster and faster while Apple made it faster and faster while leaving a few strange cracks behind. leaving a few strange cracks behind. leaving a few strange cracks behind. Slow base storage, weird monitor Slow base storage, weird monitor Slow base storage, weird monitor restrictions, and now local AI restrictions, and now local AI restrictions, and now local AI performance that's a lot more performance that's a lot more performance that's a lot more complicated than Apple's marketing makes complicated than Apple's marketing makes complicated than Apple's marketing makes it sound. The M5 appears to fix almost it sound. The M5 appears to fix almost it sound. The M5 appears to fix almost all of that. But Apple changed one more all of that. But Apple changed one more all of that. But Apple changed one more thing that might matter more to you than thing that might matter more to you than thing that might matter more to you than the development related test that I'm the development related test that I'm the development related test that I'm about to show you. Now, just so that about to show you. Now, just so that about to show you. Now, just so that we're on the same page, they're all we're on the same page, they're all we're on the same page, they're all updated to Tahoe now. The latest version updated to Tahoe now. The latest version updated to Tahoe now. The latest version as of right now. And what's incredible as of right now. And what's incredible as of right now. And what's incredible is that even the old MacBook Air M1 runs is that even the old MacBook Air M1 runs is that even the old MacBook Air M1 runs Tahoe just fine. Now, I'm not going to Tahoe just fine. Now, I'm not going to Tahoe just fine. Now, I'm not going to be glazing Apple here in this video. be glazing Apple here in this video. be glazing Apple here in this video. They're not sponsoring this. I bought They're not sponsoring this. I bought They're not sponsoring this. I bought all these myself over the last few all these myself over the last few all these myself over the last few years. So, I will be pointing out a few years. So, I will be pointing out a few years. So, I will be pointing out a few cracks here and there, but the M1 is cracks here and there, but the M1 is cracks here and there, but the M1 is still a very impressive machine, and you still a very impressive machine, and you still a very impressive machine, and you will see. Also, I'm not going to be will see. Also, I'm not going to be will see. Also, I'm not going to be comparing memory sizes because M1, M2, comparing memory sizes because M1, M2, comparing memory sizes because M1, M2, M3 all had a minimum of 8 GB of memory. M3 all had a minimum of 8 GB of memory. M3 all had a minimum of 8 GB of memory. M4, and M5 now have a minimum of 16. So, M4, and M5 now have a minimum of 16. So, M4, and M5 now have a minimum of 16. So, most of my tests are designed to measure most of my tests are designed to measure most of my tests are designed to measure the chip without turning this into an 8 the chip without turning this into an 8 the chip without turning this into an 8 versus 16 comparison. Every repository, versus 16 comparison. Every repository, versus 16 comparison. Every repository, trial account, and developer tool adds trial account, and developer tool adds trial account, and developer tool adds to your online footprint. Data brokers to your online footprint. Data brokers to your online footprint. Data brokers connect those pieces. They package them connect those pieces. They package them connect those pieces. They package them up and sell them. And less information up and sell them. And less information up and sell them. And less information about you floating out there means fewer about you floating out there means fewer about you floating out there means fewer ways they can target you. across dozens ways they can target you. across dozens ways they can target you. across dozens of countries. You can require data of countries. You can require data of countries. You can require data brokers to erase your information. But brokers to erase your information. But brokers to erase your information. But doing it yourself means finding each doing it yourself means finding each doing it yourself means finding each broker, navigating its process, and broker, navigating its process, and broker, navigating its process, and repeating it when your data returns.

repeating it when your data returns. repeating it when your data returns. Incogn sends removal requests Incogn sends removal requests Incogn sends removal requests automatically and keeps following up. automatically and keeps following up. automatically and keeps following up. Here it found 190 total brokers holding Here it found 190 total brokers holding Here it found 190 total brokers holding my details, and most have already been my details, and most have already been my details, and most have already been cleared. And with custom removals, I cleared. And with custom removals, I cleared. And with custom removals, I submit a page exposing my information, submit a page exposing my information, submit a page exposing my information, and a privacy expert handles the takeown and a privacy expert handles the takeown and a privacy expert handles the takeown and follow-ups for me. Incogn's data and follow-ups for me. Incogn's data and follow-ups for me. Incogn's data removal processes were actually verified removal processes were actually verified removal processes were actually verified by Deote. Now, it can't erase government by Deote. Now, it can't erase government by Deote. Now, it can't erase government records or random social posts, but it records or random social posts, but it records or random social posts, but it can address broker listings and eligible can address broker listings and eligible can address broker listings and eligible pages. There's also a 30-day money back pages. There's also a 30-day money back pages. There's also a 30-day money back guarantee. So, make your information guarantee. So, make your information guarantee. So, make your information harder to find with Incogn. Visit harder to find with Incogn. Visit harder to find with Incogn. Visit incogn.com/allexiscin incogn.com/allexiscin incogn.com/allexiscin and use my code alexiskin for 60% off an and use my code alexiskin for 60% off an and use my code alexiskin for 60% off an annual plan. Link is down below. annual plan. Link is down below. annual plan. Link is down below. [music] Now, from the outside, this [music] Now, from the outside, this [music] Now, from the outside, this could be just an M3 or an M2 or an M4, could be just an M3 or an M2 or an M4, could be just an M3 or an M2 or an M4, but it's an M5. Apple has apparently but it's an M5. Apple has apparently but it's an M5. Apple has apparently decided to keep the same exact chassis. decided to keep the same exact chassis. decided to keep the same exact chassis. They thought, "Hey, we did a good job. They thought, "Hey, we did a good job. They thought, "Hey, we did a good job. Uh, we're done with that chassis." Good Uh, we're done with that chassis." Good Uh, we're done with that chassis." Good job, Apple. Everything, the chassis, the job, Apple. Everything, the chassis, the job, Apple. Everything, the chassis, the keyboard, the screen, Mag Safe, and keyboard, the screen, Mag Safe, and keyboard, the screen, Mag Safe, and physical ports are all unchanged. Even physical ports are all unchanged. Even physical ports are all unchanged. Even though I really wanted Thunderbolt 5, it though I really wanted Thunderbolt 5, it though I really wanted Thunderbolt 5, it still has Thunderbolt 4 on it and still still has Thunderbolt 4 on it and still still has Thunderbolt 4 on it and still just two ports on the left side, but just two ports on the left side, but just two ports on the left side, but they gave us Wi-Fi 7. That's the big they gave us Wi-Fi 7. That's the big they gave us Wi-Fi 7. That's the big one. It's supposed to be able to move one. It's supposed to be able to move one. It's supposed to be able to move data twice as fast as old Wi-Fi. There's data twice as fast as old Wi-Fi. There's data twice as fast as old Wi-Fi. There's a new chip in it called the N1. Does it a new chip in it called the N1. Does it a new chip in it called the N1. Does it actually enable MLO and 320 MHz actually enable MLO and 320 MHz actually enable MLO and 320 MHz channels? There's also new Bluetooth on channels? There's also new Bluetooth on channels? There's also new Bluetooth on there that can tell you how far your there that can tell you how far your there that can tell you how far your devices are. It's not really my thing,

devices are. It's not really my thing, devices are. It's not really my thing, but maybe you like that. Then there you but maybe you like that. Then there you but maybe you like that. Then there you go. go. go. [music] Now, since I'm in a new place, I [music] Now, since I'm in a new place, I [music] Now, since I'm in a new place, I have a Wi-Fi 7 here and 10G wired have a Wi-Fi 7 here and 10G wired have a Wi-Fi 7 here and 10G wired network. So, I ran Iperf 3 on all these network. So, I ran Iperf 3 on all these network. So, I ran Iperf 3 on all these to see how we do. Now, before I did this to see how we do. Now, before I did this to see how we do. Now, before I did this test, I ran this on the 5 GHz network. test, I ran this on the 5 GHz network. test, I ran this on the 5 GHz network. And the M5, even though it got the And the M5, even though it got the And the M5, even though it got the highest speed, it was only 625 highest speed, it was only 625 highest speed, it was only 625 megabit per second, which is uh not megabit per second, which is uh not megabit per second, which is uh not great. Then I enabled 6 GHz, and I got great. Then I enabled 6 GHz, and I got great. Then I enabled 6 GHz, and I got these results. Yeah, not great. it lands these results. Yeah, not great. it lands these results. Yeah, not great. it lands about where the M4 is and actually a about where the M4 is and actually a about where the M4 is and actually a little bit behind the M3. That's nowhere little bit behind the M3. That's nowhere little bit behind the M3. That's nowhere near where Wi-Fi 7 is supposed to be. near where Wi-Fi 7 is supposed to be. near where Wi-Fi 7 is supposed to be. That's the real deal. And I tried That's the real deal. And I tried That's the real deal. And I tried several times and it turns out that several times and it turns out that several times and it turns out that Apple actually limits this feature to Apple actually limits this feature to Apple actually limits this feature to 160 MHz instead of giving it the whole 160 MHz instead of giving it the whole 160 MHz instead of giving it the whole 320. It's a kind of a built-in 320. It's a kind of a built-in 320. It's a kind of a built-in bottleneck. I don't know if it's because bottleneck. I don't know if it's because bottleneck. I don't know if it's because it's a special MacBook Air and not a it's a special MacBook Air and not a it's a special MacBook Air and not a MacBook Pro. TBD kind of already have a MacBook Pro. TBD kind of already have a MacBook Pro. TBD kind of already have a little bit of a loss. Don't worry, we're little bit of a loss. Don't worry, we're little bit of a loss. Don't worry, we're going to make up for it. going to make up for it. going to make up for it. record on external monitors. I tested record on external monitors. I tested record on external monitors. I tested the progression in my M4 video. The M1 the progression in my M4 video. The M1 the progression in my M4 video. The M1 and the M2 could only drive one external and the M2 could only drive one external and the M2 could only drive one external display. M3 finally gave us two, but display. M3 finally gave us two, but display. M3 finally gave us two, but only with the lid closed. And the M4 only with the lid closed. And the M4 only with the lid closed. And the M4 fixed it for real. They gave us two fixed it for real. They gave us two fixed it for real. They gave us two external displays plus the laptop screen external displays plus the laptop screen external displays plus the laptop screen all at once for a total of three. The all at once for a total of three. The all at once for a total of three. The M5, same deal as the M4. Two external M5, same deal as the M4. Two external M5, same deal as the M4. Two external displays plus the built-in display. No displays plus the built-in display. No displays plus the built-in display. No third monitor when the lid is closed.

third monitor when the lid is closed. third monitor when the lid is closed. But there is one new trick. The M5 can But there is one new trick. The M5 can But there is one new trick. The M5 can drive both external displays through a drive both external displays through a drive both external displays through a single Thunderbolt port, whereas the M4 single Thunderbolt port, whereas the M4 single Thunderbolt port, whereas the M4 had to use both ports up. That matters had to use both ports up. That matters had to use both ports up. That matters when Apple only gives you two ports. when Apple only gives you two ports. when Apple only gives you two ports. Keeping one free, that's a good thing. Keeping one free, that's a good thing. Keeping one free, that's a good thing. We're moving in the right direction We're moving in the right direction We're moving in the right direction here. [music] here. [music] here. [music] Now, four generations, Apple kept making Now, four generations, Apple kept making Now, four generations, Apple kept making the processor faster and faster while the processor faster and faster while the processor faster and faster while storage basically stood still. Boom. storage basically stood still. Boom. storage basically stood still. Boom. [music] Wow. the M1 still kicking. Even the M4, Wow. the M1 still kicking. Even the M4, the M1 has way better right speeds on the M1 has way better right speeds on the M1 has way better right speeds on that sequential. And by the way, this M1 that sequential. And by the way, this M1 that sequential. And by the way, this M1 has 512 gig SSD on it, and those are has 512 gig SSD on it, and those are has 512 gig SSD on it, and those are actually faster than the 256. These are actually faster than the 256. These are actually faster than the 256. These are 256 drives with the M2 and the M3. And 256 drives with the M2 and the M3. And 256 drives with the M2 and the M3. And let's take a look at the M5 here. Wow. let's take a look at the M5 here. Wow. let's take a look at the M5 here. Wow. Okay, this is in a different league Okay, this is in a different league Okay, this is in a different league altogether. The sequential is just altogether. The sequential is just altogether. The sequential is just insane here. And even the random is insane here. And even the random is insane here. And even the random is faster. The M5 fixed the storage problem faster. The M5 fixed the storage problem faster. The M5 fixed the storage problem so aggressively that my external drive so aggressively that my external drive so aggressively that my external drive is now the slow one. It more than is now the slow one. It more than is now the slow one. It more than doubles every machine here. Also, the M5 doubles every machine here. Also, the M5 doubles every machine here. Also, the M5 gets rid of the 256 GB model, which is gets rid of the 256 GB model, which is gets rid of the 256 GB model, which is good. I kept running out of space just good. I kept running out of space just good. I kept running out of space just installing software for testing. So, I'd installing software for testing. So, I'd installing software for testing. So, I'd imagine what real people will do. Not imagine what real people will do. Not imagine what real people will do. Not that I'm not a real person. I'm a real that I'm not a real person. I'm a real that I'm not a real person. I'm a real person. Okay. Anyway, they go up to 4 person. Okay. Anyway, they go up to 4 person. Okay. Anyway, they go up to 4 terb now, which is a nice welcome terb now, which is a nice welcome terb now, which is a nice welcome change, but that's going to cost you a change, but that's going to cost you a change, but that's going to cost you a little bit.

little bit. little bit. speedometer. It runs web applications speedometer. It runs web applications speedometer. It runs web applications written with different frameworks. It's written with different frameworks. It's written with different frameworks. It's not literally a JavaScript single core not literally a JavaScript single core not literally a JavaScript single core benchmark, but it is a good measurement benchmark, but it is a good measurement benchmark, but it is a good measurement of how responsive a complex web browser of how responsive a complex web browser of how responsive a complex web browser application would feel on the main application would feel on the main application would feel on the main thread. And this affects more than just thread. And this affects more than just thread. And this affects more than just browsing. If you build front-end browsing. If you build front-end browsing. If you build front-end applications, this is the environment applications, this is the environment applications, this is the environment you pretty much stare at all day while you pretty much stare at all day while you pretty much stare at all day while you write, you reload, you inspect, you you write, you reload, you inspect, you you write, you reload, you inspect, you debug if you're doing web apps, but also debug if you're doing web apps, but also debug if you're doing web apps, but also if you're using editors that are built if you're using editors that are built if you're using editors that are built on Electron, which if you're using VS on Electron, which if you're using VS on Electron, which if you're using VS Code, Yeah. Now, overall, the M5 is 5% Code, Yeah. Now, overall, the M5 is 5% Code, Yeah. Now, overall, the M5 is 5% over the M4. It's the smallest over the M4. It's the smallest over the M4. It's the smallest generational improvement since the M1. generational improvement since the M1. generational improvement since the M1. M1 to M2, I'm looking over there because M1 to M2, I'm looking over there because M1 to M2, I'm looking over there because that's where my chart is. I'm just that's where my chart is. I'm just that's where my chart is. I'm just setting this place up, so not everything setting this place up, so not everything setting this place up, so not everything is in the right place yet. M1 to M2 was is in the right place yet. M1 to M2 was is in the right place yet. M1 to M2 was 22%, M2 to M3 was 14% bump. M3 to M4 22%, M2 to M3 was 14% bump. M3 to M4 22%, M2 to M3 was 14% bump. M3 to M4 11%. We're getting slower and slower as 11%. We're getting slower and slower as 11%. We're getting slower and slower as far as that uh improvement here in far as that uh improvement here in far as that uh improvement here in speedometer, but as predicted, it's speedometer, but as predicted, it's speedometer, but as predicted, it's visible, but it's not huge. However, if visible, but it's not huge. However, if visible, but it's not huge. However, if you take a look at the overall you take a look at the overall you take a look at the overall improvement all the way from M1 to M5, improvement all the way from M1 to M5, improvement all the way from M1 to M5, then we're about 60% more. However, this then we're about 60% more. However, this then we're about 60% more. However, this is not the biggest difference when it is not the biggest difference when it is not the biggest difference when it comes to CPU power on the new M5. The comes to CPU power on the new M5. The comes to CPU power on the new M5. The first three machines had only eight first three machines had only eight first three machines had only eight cores. The M4 and the M5 have 10 cores, cores. The M4 and the M5 have 10 cores, cores. The M4 and the M5 have 10 cores, and most of the stuff we're going to be and most of the stuff we're going to be and most of the stuff we're going to be doing as developers, like building code, doing as developers, like building code, doing as developers, like building code, compiling code, or running compiling code, or running compiling code, or running multi-threaded applications, that's multi-threaded applications, that's multi-threaded applications, that's going to use multiple cores. But also, going to use multiple cores. But also, going to use multiple cores. But also, the improvement is coming from changed the improvement is coming from changed the improvement is coming from changed architecture. They're called super cores architecture. They're called super cores architecture. They're called super cores now in the M5. Well, at least cores 7 now in the M5. Well, at least cores 7 now in the M5. Well, at least cores 7 through 10 are called super cores.

through 10 are called super cores. through 10 are called super cores. Here's one more test. This is a C++ Here's one more test. This is a C++ Here's one more test. This is a C++ compiled application, but using only one compiled application, but using only one compiled application, but using only one core. And the way it optimizes the code core. And the way it optimizes the code core. And the way it optimizes the code as it's running is it bounces around the as it's running is it bounces around the as it's running is it bounces around the available cores, handing things off available cores, handing things off available cores, handing things off dynamically from core to core. And there dynamically from core to core. And there dynamically from core to core. And there it is running on the M1 using those it is running on the M1 using those it is running on the M1 using those performance cores. That's what it looks performance cores. That's what it looks performance cores. That's what it looks like on the M2. also have eight like on the M2. also have eight like on the M2. also have eight processor cores, two here, and we've got processor cores, two here, and we've got processor cores, two here, and we've got the sort happening on those performance the sort happening on those performance the sort happening on those performance cores. Well, on one at a time. It's kind cores. Well, on one at a time. It's kind cores. Well, on one at a time. It's kind of bouncing between those. Same thing of bouncing between those. Same thing of bouncing between those. Same thing here on the M3. On the M4, we have now here on the M3. On the M4, we have now here on the M3. On the M4, we have now 10 cores and also 10 cores on the M5. 10 cores and also 10 cores on the M5. 10 cores and also 10 cores on the M5. The result, 2 minutes and 38 seconds, 2 The result, 2 minutes and 38 seconds, 2 The result, 2 minutes and 38 seconds, 2 minutes and 42 seconds, a little bit minutes and 42 seconds, a little bit minutes and 42 seconds, a little bit worse on the M2. 2 minutes 57 seconds, worse on the M2. 2 minutes 57 seconds, worse on the M2. 2 minutes 57 seconds, better on the M3, 2 minutes 31 seconds, better on the M3, 2 minutes 31 seconds, better on the M3, 2 minutes 31 seconds, better yet. And finally, 2 minutes and better yet. And finally, 2 minutes and better yet. And finally, 2 minutes and 19 seconds on the M5. It's all 19 seconds on the M5. It's all 19 seconds on the M5. It's all incremental, but we're getting better incremental, but we're getting better incremental, but we're getting better and better. And the M5 is about 8% and better. And the M5 is about 8% and better. And the M5 is about 8% faster than the M4. That's just pure faster than the M4. That's just pure faster than the M4. That's just pure gain since core count didn't change from gain since core count didn't change from gain since core count didn't change from M4 to M5. Now, just for giggles, the M2 M4 to M5. Now, just for giggles, the M2 M4 to M5. Now, just for giggles, the M2 actually came back a little bit slower actually came back a little bit slower actually came back a little bit slower than the M1. I know that's within the than the M1. I know that's within the than the M1. I know that's within the margin of error, but still, that was a margin of error, but still, that was a margin of error, but still, that was a bad year, I think. Big picture, M1 to M5 bad year, I think. Big picture, M1 to M5 bad year, I think. Big picture, M1 to M5 cut single core sort time by about 36%.

cut single core sort time by about 36%. cut single core sort time by about 36%. Now, this is the kind of improvement Now, this is the kind of improvement Now, this is the kind of improvement that you feel everywhere, but it may not that you feel everywhere, but it may not that you feel everywhere, but it may not be the kind that justifies replacing the be the kind that justifies replacing the be the kind that justifies replacing the M4. There is something else that might. M4. There is something else that might. M4. There is something else that might. [music] [music] [music] Sorting integers is fun, but developers Sorting integers is fun, but developers Sorting integers is fun, but developers don't buy new laptops to sort things. don't buy new laptops to sort things. don't buy new laptops to sort things. This time, I'm going to do an actual This time, I'm going to do an actual This time, I'm going to do an actual build. I used a fixed llama CPP commit. build. I used a fixed llama CPP commit. build. I used a fixed llama CPP commit. Yeah, that Llama CPP, the famous machine Yeah, that Llama CPP, the famous machine Yeah, that Llama CPP, the famous machine learning library or tool, whatever you learning library or tool, whatever you learning library or tool, whatever you want to call it. Llama CPP build and want to call it. Llama CPP build and want to call it. Llama CPP build and it's complete. M1 1 minute 40 seconds to it's complete. M1 1 minute 40 seconds to it's complete. M1 1 minute 40 seconds to build that. M2 1 minute 22 seconds. M3 1 build that. M2 1 minute 22 seconds. M3 1 build that. M2 1 minute 22 seconds. M3 1 minute 14 seconds. Getting faster. M4 1 minute 14 seconds. Getting faster. M4 1 minute 14 seconds. Getting faster. M4 1 minute 3 seconds and breaking 1 minute minute 3 seconds and breaking 1 minute minute 3 seconds and breaking 1 minute 54 seconds on the M5. And it's about 14% 54 seconds on the M5. And it's about 14% 54 seconds on the M5. And it's about 14% faster than the M4. It cuts the M1's faster than the M4. It cuts the M1's faster than the M4. It cuts the M1's time nearly in half. We're talking about time nearly in half. We're talking about time nearly in half. We're talking about real gains here. Every generation showed real gains here. Every generation showed real gains here. Every generation showed us better improvements with the us better improvements with the us better improvements with the multi-core throughput. And this is not multi-core throughput. And this is not multi-core throughput. And this is not just the CPU cores. This is everything. just the CPU cores. This is everything. just the CPU cores. This is everything. Everything is involved here. The SSD, Everything is involved here. The SSD, Everything is involved here. The SSD, the CPU cores, the memory, everything.

the CPU cores, the memory, everything. the CPU cores, the memory, everything. But there is a little bit of concern But there is a little bit of concern But there is a little bit of concern from people that think that these from people that think that these from people that think that these machines wo that these machines machines wo that these machines machines wo that these machines throttle. And last year's M4 did get throttle. And last year's M4 did get throttle. And last year's M4 did get pretty warm, a lot warmer than the pretty warm, a lot warmer than the pretty warm, a lot warmer than the previous models. So the M5, is it going previous models. So the M5, is it going previous models. So the M5, is it going to get even warmer now? And is it going to get even warmer now? And is it going to get even warmer now? And is it going to throttle when we throw really hard to throttle when we throw really hard to throttle when we throw really hard work at it? I had a handy benchmark. By work at it? I had a handy benchmark. By work at it? I had a handy benchmark. By the way, my benchmarks are available on the way, my benchmarks are available on the way, my benchmarks are available on my GitHub. Oh, I have a handy C++ my GitHub. Oh, I have a handy C++ my GitHub. Oh, I have a handy C++ program. It's a merge sort and it uses program. It's a merge sort and it uses program. It's a merge sort and it uses all the cores. Yeah, again, it's a sort, all the cores. Yeah, again, it's a sort, all the cores. Yeah, again, it's a sort, I know, but this one is a serious one. I know, but this one is a serious one. I know, but this one is a serious one. It sorts 1 billion integers using all It sorts 1 billion integers using all It sorts 1 billion integers using all the cores possible and different the cores possible and different the cores possible and different algorithm. Merge sort. Actually, it's my algorithm. Merge sort. Actually, it's my algorithm. Merge sort. Actually, it's my favorite sort. I love merge sort. I favorite sort. I love merge sort. I favorite sort. I love merge sort. I know. Call me a nerd in the comments. know. Call me a nerd in the comments. know. Call me a nerd in the comments. It's fine. It's fine. I'm used to it. It's fine. It's fine. I'm used to it. It's fine. It's fine. I'm used to it. All right. I'm got all these set up. All right. I'm got all these set up. All right. I'm got all these set up. Doing a sustained test. And of course, I got my old friend here. And of course, I got my old friend here. For those of you that know what this is, For those of you that know what this is, For those of you that know what this is, it's been a while. Comment down below. it's been a while. Comment down below. it's been a while. Comment down below. It's been a while, old friend. By the It's been a while, old friend. By the It's been a while, old friend. By the way, this is going to be a loop running way, this is going to be a loop running way, this is going to be a loop running merge sort on a billion integers and merge sort on a billion integers and merge sort on a billion integers and using up all the cores. So, I expect the using up all the cores. So, I expect the using up all the cores. So, I expect the two machines that have 10 cores to two machines that have 10 cores to two machines that have 10 cores to obviously win. But which one of them is obviously win. But which one of them is obviously win. But which one of them is going to win? And which one of them is going to win? And which one of them is going to win? And which one of them is not going to throttle? We're about to not going to throttle? We're about to not going to throttle? We're about to find out.

find out. find out. [laughter] [laughter] [laughter] There they go. And you can see that this There they go. And you can see that this There they go. And you can see that this is definitely using up all the cores. is definitely using up all the cores. is definitely using up all the cores. All those green soldiers marching on. All those green soldiers marching on. All those green soldiers marching on. There's the M4, M3. They're all very, There's the M4, M3. They're all very, There's the M4, M3. They're all very, very busy right now. M1 also pretty very busy right now. M1 also pretty very busy right now. M1 also pretty busy. Iteration one is done on the M3, busy. Iteration one is done on the M3, busy. Iteration one is done on the M3, M4, [snorts] and M5. You can see where M4, [snorts] and M5. You can see where M4, [snorts] and M5. You can see where that dipped down a little bit and kicked that dipped down a little bit and kicked that dipped down a little bit and kicked right back up for iteration two. We're right back up for iteration two. We're right back up for iteration two. We're doing five iterations. And uh just doing five iterations. And uh just doing five iterations. And uh just finished with iteration one on the M1 finished with iteration one on the M1 finished with iteration one on the M1 and the M2. all going according to plan. and the M2. all going according to plan. and the M2. all going according to plan. Now, if you watch my video on the M4 Now, if you watch my video on the M4 Now, if you watch my video on the M4 last year, you know that this machine last year, you know that this machine last year, you know that this machine got pretty warm. The palm rests are got pretty warm. The palm rests are got pretty warm. The palm rests are warm. They're not uncomfortable. They're warm. They're not uncomfortable. They're warm. They're not uncomfortable. They're just warm. They're like, it feels like I just warm. They're like, it feels like I just warm. They're like, it feels like I don't know, feels very warm, but it's don't know, feels very warm, but it's don't know, feels very warm, but it's not like hot. This not like hot. This not like hot. This [laughter] [laughter] [laughter] uh it feels about the same. Maybe just a uh it feels about the same. Maybe just a uh it feels about the same. Maybe just a tad bit warmer for the M5. M3 feels tad bit warmer for the M5. M3 feels tad bit warmer for the M5. M3 feels warm, M2 feels cooler, and M1 feels the warm, M2 feels cooler, and M1 feels the warm, M2 feels cooler, and M1 feels the coolest. How did they get the M1 so coolest. How did they get the M1 so coolest. How did they get the M1 so right? For those of you that are right? For those of you that are right? For those of you that are wondering what I'm doing here, this wondering what I'm doing here, this wondering what I'm doing here, this crazy amount of processing requires a crazy amount of processing requires a crazy amount of processing requires a lot of power and in a small body without lot of power and in a small body without lot of power and in a small body without cooling, it might just have to slow that cooling, it might just have to slow that cooling, it might just have to slow that CPU down in order to keep up and not CPU down in order to keep up and not CPU down in order to keep up and not overheat. Usually in a MacBook Pro, for overheat. Usually in a MacBook Pro, for overheat. Usually in a MacBook Pro, for example, you'd have fans going so it example, you'd have fans going so it example, you'd have fans going so it doesn't slow down even under high load doesn't slow down even under high load doesn't slow down even under high load like this. But on MacBook Airs, it might like this. But on MacBook Airs, it might like this. But on MacBook Airs, it might slow down. That's why we're pushing it slow down. That's why we're pushing it slow down. That's why we're pushing it really hard to see what happens. The really hard to see what happens. The really hard to see what happens. The funny thing is that M1 is actually funny thing is that M1 is actually funny thing is that M1 is actually finishing up iteration two while the M2

finishing up iteration two while the M2 finishing up iteration two while the M2 is still working on finishing up is still working on finishing up is still working on finishing up iteration one. Oh, it just started iteration one. Oh, it just started iteration one. Oh, it just started iteration two. And we have a winner. iteration two. And we have a winner. iteration two. And we have a winner. It's the uh it's the M5. Finished five It's the uh it's the M5. Finished five It's the uh it's the M5. Finished five iterations. 11 minutes and 33 seconds. iterations. 11 minutes and 33 seconds. iterations. 11 minutes and 33 seconds. All the other machines are still working All the other machines are still working All the other machines are still working at it. Some machines are on the third at it. Some machines are on the third at it. Some machines are on the third iteration. But I do have to wonder iteration. But I do have to wonder iteration. But I do have to wonder between the M3 and the M4 which one of between the M3 and the M4 which one of between the M3 and the M4 which one of them is going to come in next. And I am them is going to come in next. And I am them is going to come in next. And I am seeing slowdowns. Thanks old friend. seeing slowdowns. Thanks old friend. seeing slowdowns. Thanks old friend. Don't need you anymore. All right, Don't need you anymore. All right, Don't need you anymore. All right, they're finally all done. And this is they're finally all done. And this is they're finally all done. And this is pretty significant because this pretty significant because this pretty significant because this represents those long running processes represents those long running processes represents those long running processes that developers that build large code that developers that build large code that developers that build large code bases will face taken to the 10th power. bases will face taken to the 10th power. bases will face taken to the 10th power. So it's a little bit extreme on a So it's a little bit extreme on a So it's a little bit extreme on a smaller scale. Let me get into the smaller scale. Let me get into the smaller scale. Let me get into the results. Okay, we've got a couple results. Okay, we've got a couple results. Okay, we've got a couple machines here that took a while. That's machines here that took a while. That's machines here that took a while. That's the M1 and the M2. M1's last iteration the M1 and the M2. M1's last iteration the M1 and the M2. M1's last iteration took 222 seconds. M26. took 222 seconds. M26. took 222 seconds. M26. Not a huge difference there. And if you Not a huge difference there. And if you Not a huge difference there. And if you examine the overall time, the M2 did the examine the overall time, the M2 did the examine the overall time, the M2 did the whole job slower than the M1. So that's whole job slower than the M1. So that's whole job slower than the M1. So that's kind of our loser as far as speed goes.

kind of our loser as far as speed goes. kind of our loser as far as speed goes. But as far as variation, it's actually But as far as variation, it's actually But as far as variation, it's actually the winner. It only had a change of 3% the winner. It only had a change of 3% the winner. It only had a change of 3% from the first iteration to the last. M3 from the first iteration to the last. M3 from the first iteration to the last. M3 and M4 195 seconds and 149 seconds. much and M4 195 seconds and 149 seconds. much and M4 195 seconds and 149 seconds. much better, but they both had a 5% cut in better, but they both had a 5% cut in better, but they both had a 5% cut in performance from iteration 1 to performance from iteration 1 to performance from iteration 1 to iteration 5, which means that each iteration 5, which means that each iteration 5, which means that each iteration took a little bit longer to do iteration took a little bit longer to do iteration took a little bit longer to do because the processor was throttled. The because the processor was throttled. The because the processor was throttled. The fastest is the M5, of course, with 129 fastest is the M5, of course, with 129 fastest is the M5, of course, with 129 seconds, but it had the largest variance seconds, but it had the largest variance seconds, but it had the largest variance of 6%. It's not much, but it's of 6%. It's not much, but it's of 6%. It's not much, but it's definitely there. So, that means from definitely there. So, that means from definitely there. So, that means from iteration 1 to iteration 5, it slowed iteration 1 to iteration 5, it slowed iteration 1 to iteration 5, it slowed down the most. All right, we have down the most. All right, we have down the most. All right, we have compiled code. Now let's give it some compiled code. Now let's give it some compiled code. Now let's give it some Python and make these machines really Python and make these machines really Python and make these machines really suffer. suffer. suffer. Now traditionally I do the Mandlerro Now traditionally I do the Mandlerro Now traditionally I do the Mandlerro test which is different than the one we test which is different than the one we test which is different than the one we just did. That was C++ compiled. This just did. That was C++ compiled. This just did. That was C++ compiled. This one is an interpreted test. It's Python one is an interpreted test. It's Python one is an interpreted test. It's Python and I know that some of you are Python and I know that some of you are Python and I know that some of you are Python developers out there. So let's go. developers out there. So let's go. developers out there. So let's go. Sometimes it's just fun to press them Sometimes it's just fun to press them Sometimes it's just fun to press them all instead of letting that guy do all all instead of letting that guy do all all instead of letting that guy do all the work. This one is also pretty the work. This one is also pretty the work. This one is also pretty intense because it pegs all the cores intense because it pegs all the cores intense because it pegs all the cores completely. I'm expecting these two completely. I'm expecting these two completely. I'm expecting these two machines, the M4 and the M5 to win machines, the M4 and the M5 to win machines, the M4 and the M5 to win again, but we'll see. Oh. Oh, that was again, but we'll see. Oh. Oh, that was again, but we'll see. Oh. Oh, that was really close. M4 and M5 finished. Then really close. M4 and M5 finished. Then really close. M4 and M5 finished. Then this one, the M3 and the M2. And this one, the M3 and the M2. And this one, the M3 and the M2. And well, well, well, finally the M1 at 56 well, well, well, finally the M1 at 56 well, well, well, finally the M1 at 56 total seconds for that one. 50 seconds total seconds for that one. 50 seconds total seconds for that one. 50 seconds for the M2, 43 on the M3, 34, and not for the M2, 43 on the M3, 34, and not for the M2, 43 on the M3, 34, and not that much faster. Only 32 seconds, about

that much faster. Only 32 seconds, about that much faster. Only 32 seconds, about 2 and 1/2 seconds faster on the M5. 2 and 1/2 seconds faster on the M5. 2 and 1/2 seconds faster on the M5. Now, it's time to change gears and talk Now, it's time to change gears and talk Now, it's time to change gears and talk about AI. Well, before we get into AI, about AI. Well, before we get into AI, about AI. Well, before we get into AI, there is the memory issue, right? there is the memory issue, right? there is the memory issue, right? Because AI uses two components, memory Because AI uses two components, memory Because AI uses two components, memory and the GPU. First, it does the and the GPU. First, it does the and the GPU. First, it does the processing and then it uses the memory processing and then it uses the memory processing and then it uses the memory bandwidth to do the token generation. bandwidth to do the token generation. bandwidth to do the token generation. And we have a test for that. The stream And we have a test for that. The stream And we have a test for that. The stream benchmark. It's a long live staple and benchmark. It's a long live staple and benchmark. It's a long live staple and it's been updated [music] to support the it's been updated [music] to support the it's been updated [music] to support the M5s. And the M5 copy is 142 GB per M5s. And the M5 copy is 142 GB per M5s. And the M5 copy is 142 GB per second. That's a 26% jump over the M4's second. That's a 26% jump over the M4's second. That's a 26% jump over the M4's 113. What's interesting is that it's 113. What's interesting is that it's 113. What's interesting is that it's almost exactly Apple's claimed 28% almost exactly Apple's claimed 28% almost exactly Apple's claimed 28% increase. And in absolute terms, it's increase. And in absolute terms, it's increase. And in absolute terms, it's the single biggest generational the single biggest generational the single biggest generational bandwidth leap in the MacBook Air lineup bandwidth leap in the MacBook Air lineup bandwidth leap in the MacBook Air lineup since the M1. You know why? AI. since the M1. You know why? AI. since the M1. You know why? AI. I'm going to stop now. AI. I'm going to stop now. AI. I'm going to stop now. AI. [music] We're not done with AI yet [music] We're not done with AI yet [music] We're not done with AI yet because local AI is a thing now and because local AI is a thing now and because local AI is a thing now and Apple has been really focusing on it.

Apple has been really focusing on it. Apple has been really focusing on it. Last time I treated local AI performance Last time I treated local AI performance Last time I treated local AI performance as one number and that kind of missed as one number and that kind of missed as one number and that kind of missed what M5 this year is actually really what M5 this year is actually really what M5 this year is actually really good at and what's changed the most. good at and what's changed the most. good at and what's changed the most. See, every M5 GPU core now contains a See, every M5 GPU core now contains a See, every M5 GPU core now contains a neural accelerator. And you might neural accelerator. And you might neural accelerator. And you might wonder, well, what the heck does that wonder, well, what the heck does that wonder, well, what the heck does that do? Well, it shows up. All right, there do? Well, it shows up. All right, there do? Well, it shows up. All right, there we go. And there is the blue marching we go. And there is the blue marching we go. And there is the blue marching man, which I call the GPU usage right man, which I call the GPU usage right man, which I call the GPU usage right here. And we've got decode happening here. And we've got decode happening here. And we've got decode happening already on the M5. Decode here. Nice already on the M5. Decode here. Nice already on the M5. Decode here. Nice tokens per second right there. Almost 30 tokens per second right there. Almost 30 tokens per second right there. Almost 30 on the M3. And we're still in the on the M3. And we're still in the on the M3. And we're still in the prefill stage on the M1 and on the M2. prefill stage on the M1 and on the M2. prefill stage on the M1 and on the M2. Oh, M2 is now decoding. And here we go. Oh, M2 is now decoding. And here we go. Oh, M2 is now decoding. And here we go. And there's the blue marching soldiers And there's the blue marching soldiers And there's the blue marching soldiers on the history charts of all these. on the history charts of all these. on the history charts of all these. Okay. Wow. Okay. Wow. Okay. Wow. This is different. Now, if you're going This is different. Now, if you're going This is different. Now, if you're going to be doing AI, there's only one machine to be doing AI, there's only one machine to be doing AI, there's only one machine here that really, really kicks it up a here that really, really kicks it up a here that really, really kicks it up a notch. Let's start from over here, shall notch. Let's start from over here, shall notch. Let's start from over here, shall we? I'll get to that. So, this is MLX we? I'll get to that. So, this is MLX we? I'll get to that. So, this is MLX Quen 34B 4bit. Yeah, it's a small model, Quen 34B 4bit. Yeah, it's a small model, Quen 34B 4bit. Yeah, it's a small model, but these are all small machines with but these are all small machines with but these are all small machines with very little memory. So, you want to fit very little memory. So, you want to fit very little memory. So, you want to fit all your models and all the context that all your models and all the context that all your models and all the context that comes with it inside the memory comes with it inside the memory comes with it inside the memory footprint. Some of these have 8 gigs of footprint. Some of these have 8 gigs of footprint. Some of these have 8 gigs of memory. Yeah. 19.3 tokens per second memory. Yeah. 19.3 tokens per second memory. Yeah. 19.3 tokens per second here on the M1. 26.6 here on the M1. 26.6 here on the M1. 26.6 on the M2, 30 tokens per second on the on the M2, 30 tokens per second on the on the M2, 30 tokens per second on the M3. We're getting faster and faster.

M3. We're getting faster and faster. M3. We're getting faster and faster. That's good. It kind of follows that That's good. It kind of follows that That's good. It kind of follows that memory bandwidth stream benchmark that memory bandwidth stream benchmark that memory bandwidth stream benchmark that we just did. 33.3 tokens per second on we just did. 33.3 tokens per second on we just did. 33.3 tokens per second on the M4 and quite a lot more. 43.9 tokens the M4 and quite a lot more. 43.9 tokens the M4 and quite a lot more. 43.9 tokens per second on the M5. What am I excited per second on the M5. What am I excited per second on the M5. What am I excited about then? Well, if you have a keen about then? Well, if you have a keen about then? Well, if you have a keen eye, you might have noticed another eye, you might have noticed another eye, you might have noticed another number here that's very important. The number here that's very important. The number here that's very important. The number I just mentioned is token number I just mentioned is token number I just mentioned is token generation. That's how many tokens per generation. That's how many tokens per generation. That's how many tokens per second we're actually putting out. But second we're actually putting out. But second we're actually putting out. But what about prompt processing? That's the what about prompt processing? That's the what about prompt processing? That's the PP number here. 135 tokens per second on PP number here. 135 tokens per second on PP number here. 135 tokens per second on the M1, 150 on the M2, 255 on the M3, the M1, 150 on the M2, 255 on the M3, the M1, 150 on the M2, 255 on the M3, 284 on the M4. And this number is 284 on the M4. And this number is 284 on the M4. And this number is important because that's the number important because that's the number important because that's the number you're going to actually use when you're you're going to actually use when you're you're going to actually use when you're having a large prompt. And if you're having a large prompt. And if you're having a large prompt. And if you're using this in a coding scenario where using this in a coding scenario where using this in a coding scenario where you're hooking it up to your code you're hooking it up to your code you're hooking it up to your code editor, this is the number that's either editor, this is the number that's either editor, this is the number that's either going to slow you down or speed you up. going to slow you down or speed you up. going to slow you down or speed you up. It's a break it or make it moment. And It's a break it or make it moment. And It's a break it or make it moment. And look at this right here. 1,61 look at this right here. 1,61 look at this right here. 1,61 tokens per second on the M5. That is tokens per second on the M5. That is tokens per second on the M5. That is huge. That's 3.7 times the M4. From 284 huge. That's 3.7 times the M4. From 284 huge. That's 3.7 times the M4. From 284 to 1,61 tokens per second, which almost to 1,61 tokens per second, which almost to 1,61 tokens per second, which almost exactly lines up with Apple's four times exactly lines up with Apple's four times exactly lines up with Apple's four times claim. And from M1 to M5, that's almost claim. And from M1 to M5, that's almost claim. And from M1 to M5, that's almost eight times preill improvement.

eight times preill improvement. eight times preill improvement. That is a result of those neural That is a result of those neural That is a result of those neural accelerators in the GPU cores. Yes, the accelerators in the GPU cores. Yes, the accelerators in the GPU cores. Yes, the memory bandwidth is faster on that, but memory bandwidth is faster on that, but memory bandwidth is faster on that, but it's incrementally faster. This is the it's incrementally faster. This is the it's incrementally faster. This is the actual GPU processing power that's going actual GPU processing power that's going actual GPU processing power that's going to dramatically improve your AI tasks. to dramatically improve your AI tasks. to dramatically improve your AI tasks. Now, the M4 launched at $999 Now, the M4 launched at $999 Now, the M4 launched at $999 with 16 GB and 256 GB. You know which with 16 GB and 256 GB. You know which with 16 GB and 256 GB. You know which one is which. The M5 launched at $1,99 one is which. The M5 launched at $1,99 one is which. The M5 launched at $1,99 with 16 GB and 512 GB. At 1,99, this was with 16 GB and 512 GB. At 1,99, this was with 16 GB and 512 GB. At 1,99, this was easy. you were paying only $100 more and easy. you were paying only $100 more and easy. you were paying only $100 more and getting twice the storage and getting twice the storage and getting twice the storage and dramatically faster SSD, faster CPU, dramatically faster SSD, faster CPU, dramatically faster SSD, faster CPU, faster GPU, better local AI performance. faster GPU, better local AI performance. faster GPU, better local AI performance. But Apple raised prices across the Mac But Apple raised prices across the Mac But Apple raised prices across the Mac and iPad lineup, and the 13-in M5 Air and iPad lineup, and the 13-in M5 Air and iPad lineup, and the 13-in M5 Air went from 1,99 went from 1,99 went from 1,99 to,299. to,299. to,299. Now, Apple stated the reason memory cost Now, Apple stated the reason memory cost Now, Apple stated the reason memory cost crisis with AI data centers absorbing crisis with AI data centers absorbing crisis with AI data centers absorbing the global supply of RAM. I don't know the global supply of RAM. I don't know the global supply of RAM. I don't know how long this is going to last, but this how long this is going to last, but this how long this is going to last, but this may be the new normal. As unfortunate as may be the new normal. As unfortunate as may be the new normal. As unfortunate as it may be, it is a beast of a machine, it may be, it is a beast of a machine, it may be, it is a beast of a machine, but if you don't need that AI but if you don't need that AI but if you don't need that AI performance, check out last year's performance, check out last year's performance, check out last year's model. Apple finally fixed the base model. Apple finally fixed the base model. Apple finally fixed the base MacBook Air, but then they added $200 to MacBook Air, but then they added $200 to MacBook Air, but then they added $200 to the price. Now, sometimes there's deals, the price. Now, sometimes there's deals, the price. Now, sometimes there's deals, and if I find any, I'll link to them and if I find any, I'll link to them and if I find any, I'll link to them down below. But the question is no down below. But the question is no down below. But the question is no longer whether the M5 is better. Of longer whether the M5 is better. Of longer whether the M5 is better. Of course, it's better, and in some cases, course, it's better, and in some cases, course, it's better, and in some cases, it's a lot better. [laughter] it's a lot better. [laughter] it's a lot better. [laughter] The question is whether these fixes are

The question is whether these fixes are The question is whether these fixes are worth the gap now between this and the worth the gap now between this and the worth the gap now between this and the discounted M4. If you already have an discounted M4. If you already have an discounted M4. If you already have an M4, keep it. Sure, the M5 has a much M4, keep it. Sure, the M5 has a much M4, keep it. Sure, the M5 has a much faster SSD and more interesting AI faster SSD and more interesting AI faster SSD and more interesting AI architecture, but it does not make the architecture, but it does not make the architecture, but it does not make the M4 obsolete. If you are lower than that, M4 obsolete. If you are lower than that, M4 obsolete. If you are lower than that, M3, M2, M1, and you're running out of M3, M2, M1, and you're running out of M3, M2, M1, and you're running out of memory or you're running out of storage, memory or you're running out of storage, memory or you're running out of storage, you need multiple external displays, you need multiple external displays, you need multiple external displays, then think about the upgrade. Otherwise, then think about the upgrade. Otherwise, then think about the upgrade. Otherwise, if you're doing ordinary development, if you're doing ordinary development, if you're doing ordinary development, probably doesn't demand an upgrade. The probably doesn't demand an upgrade. The probably doesn't demand an upgrade. The M1 still pretty surprising to this day, M1 still pretty surprising to this day, M1 still pretty surprising to this day, 2026. And ever since then, Apple has 2026. And ever since then, Apple has 2026. And ever since then, Apple has been sealing the cracks one by one. And been sealing the cracks one by one. And been sealing the cracks one by one. And now that they got to the M5, it seems now that they got to the M5, it seems now that they got to the M5, it seems like they're almost done sealing the like they're almost done sealing the like they're almost done sealing the cracks, except for the one in your cracks, except for the one in your cracks, except for the one in your wallet. They made that one a little bit wallet. They made that one a little bit wallet. They made that one a little bit wider. If you want to watch my M4 video wider. If you want to watch my M4 video wider. If you want to watch my M4 video for devs, watch that over here. Thanks for devs, watch that over here. Thanks for devs, watch that over here. Thanks for watching and I'll see you next time.