I Waited Two Years for This Mini PC… Then I Tested It

Two years ago Qualcomm made this. This Two years ago Qualcomm made this. This is the Snapdragon X Elite dev kit. Only is the Snapdragon X Elite dev kit. Only is the Snapdragon X Elite dev kit. Only a handful of these exist in the world a handful of these exist in the world a handful of these exist in the world because they canceled it. It does look because they canceled it. It does look because they canceled it. It does look pretty cool though. Windows on ARM with pretty cool though. Windows on ARM with pretty cool though. Windows on ARM with the X Elite chip never got its desktop the X Elite chip never got its desktop the X Elite chip never got its desktop until now. So, what's the fastest until now. So, what's the fastest until now. So, what's the fastest Windows on ARM desktop you can actually Windows on ARM desktop you can actually Windows on ARM desktop you can actually buy? It's this thing. And what's the buy? It's this thing. And what's the buy? It's this thing. And what's the first mini PC with a flagship Snapdragon first mini PC with a flagship Snapdragon first mini PC with a flagship Snapdragon chip in it that isn't a dev kit and chip in it that isn't a dev kit and chip in it that isn't a dev kit and didn't get canceled? Yeah, it's also didn't get canceled? Yeah, it's also didn't get canceled? Yeah, it's also that one. So, this is the Mac mini that one. So, this is the Mac mini that one. So, this is the Mac mini killer, right? Well, I have no idea. killer, right? Well, I have no idea. killer, right? Well, I have no idea. That's what we're going to do today. That's what we're going to do today. That's what we're going to do today. Let's go. So, this is the Asus S and Q Let's go. So, this is the Asus S and Q Let's go. So, this is the Asus S and Q and 10. It's the first mini PC ever with and 10. It's the first mini PC ever with and 10. It's the first mini PC ever with a Snapdragon X2 Elite inside. Mine has a Snapdragon X2 Elite inside. Mine has a Snapdragon X2 Elite inside. Mine has 32 gigs of memory and half a terabyte of 32 gigs of memory and half a terabyte of 32 gigs of memory and half a terabyte of storage, three USB four ports on the storage, three USB four ports on the storage, three USB four ports on the back, two and a half gig ethernet. Uh, back, two and a half gig ethernet. Uh, back, two and a half gig ethernet. Uh, it could be a little faster these days. it could be a little faster these days. it could be a little faster these days. It's 2026. Let's get up to 10 already. It's 2026. Let's get up to 10 already. It's 2026. Let's get up to 10 already. And it's got Wi-Fi 7. And yes, it does And it's got Wi-Fi 7. And yes, it does And it's got Wi-Fi 7. And yes, it does have a power brick, 180 watts. The Mac have a power brick, 180 watts. The Mac have a power brick, 180 watts. The Mac mini doesn't. Come on everybody, it's mini doesn't. Come on everybody, it's mini doesn't. Come on everybody, it's 2026.

2026. 2026. Do I have to keep saying that? Also, is Do I have to keep saying that? Also, is Do I have to keep saying that? Also, is it going to use all that 180 watts? it going to use all that 180 watts? it going to use all that 180 watts? It's supposed to be a very efficient It's supposed to be a very efficient It's supposed to be a very efficient chip, so we'll see. Speaking of that chip, so we'll see. Speaking of that chip, so we'll see. Speaking of that chip, the X2 Elite, this is pretty much chip, the X2 Elite, this is pretty much chip, the X2 Elite, this is pretty much the reason this box exists. The X2 Elite the reason this box exists. The X2 Elite the reason this box exists. The X2 Elite has third generation Orion cores, 3 has third generation Orion cores, 3 has third generation Orion cores, 3 nanometers. They were announced last nanometers. They were announced last nanometers. They were announced last September. And until this, every single September. And until this, every single September. And until this, every single X2 Elite machine has been a laptop. Here X2 Elite machine has been a laptop. Here X2 Elite machine has been a laptop. Here are the X2 Elites listed on the are the X2 Elites listed on the are the X2 Elites listed on the Qualcomm's website. We're looking at the Qualcomm's website. We're looking at the Qualcomm's website. We're looking at the X2E, X2 Elite, 88100. And we've got 18 X2E, X2 Elite, 88100. And we've got 18 X2E, X2 Elite, 88100. And we've got 18 cores, 12 prime and six performance cores, 12 prime and six performance cores, 12 prime and six performance cores. They're all big cores, no little cores. They're all big cores, no little cores. They're all big cores, no little efficiency cores like Apple and Intel efficiency cores like Apple and Intel efficiency cores like Apple and Intel do. Notice the first generation X Elite do. Notice the first generation X Elite do. Notice the first generation X Elite chips topped out at just 12 cores. Some chips topped out at just 12 cores. Some chips topped out at just 12 cores. Some nerdy specs here, up to 4.7 GHz boost. nerdy specs here, up to 4.7 GHz boost. nerdy specs here, up to 4.7 GHz boost. This is not the one that made the This is not the one that made the This is not the one that made the headlines being the first 5 GHz ARM headlines being the first 5 GHz ARM headlines being the first 5 GHz ARM chip. That would be the X2 Elite chip. That would be the X2 Elite chip. That would be the X2 Elite Extreme. So, not this one. This one does Extreme. So, not this one. This one does Extreme. So, not this one. This one does have a wide memory bus LPDDR5X, have a wide memory bus LPDDR5X, have a wide memory bus LPDDR5X, although the memory bus is not as wide although the memory bus is not as wide although the memory bus is not as wide as the extreme box. That one goes up to as the extreme box. That one goes up to as the extreme box. That one goes up to 228 GB per second of memory bandwidth.

228 GB per second of memory bandwidth. 228 GB per second of memory bandwidth. This one goes up to 152. The M4 Mac Mini This one goes up to 152. The M4 Mac Mini This one goes up to 152. The M4 Mac Mini has 120 GB per second memory bandwidth, has 120 GB per second memory bandwidth, has 120 GB per second memory bandwidth, and the M4 Pro 273. It's got a new GPU and the M4 Pro 273. It's got a new GPU and the M4 Pro 273. It's got a new GPU in there. The Adreno X290. Qualcomm says in there. The Adreno X290. Qualcomm says in there. The Adreno X290. Qualcomm says it's 2.3 times the performance per watt it's 2.3 times the performance per watt it's 2.3 times the performance per watt from the first generation. And that from the first generation. And that from the first generation. And that matters for AI tests also because the matters for AI tests also because the matters for AI tests also because the GPU is responsible for the actual matrix GPU is responsible for the actual matrix GPU is responsible for the actual matrix multiplication, so the calculations. multiplication, so the calculations. multiplication, so the calculations. That's the first part of inference. And That's the first part of inference. And That's the first part of inference. And the NPU, Hexagon 80 tops, up from 45. the NPU, Hexagon 80 tops, up from 45. the NPU, Hexagon 80 tops, up from 45. Numbers look good. Now, why should we, Numbers look good. Now, why should we, Numbers look good. Now, why should we, as developers, care about this? Visual as developers, care about this? Visual as developers, care about this? Visual Studio, .NET, Node, Python, they're all Studio, .NET, Node, Python, they're all Studio, .NET, Node, Python, they're all native arm64 now. But the question native arm64 now. But the question native arm64 now. But the question remains, is it fast? I don't expect one remains, is it fast? I don't expect one remains, is it fast? I don't expect one AI prompt to build an entire [music] AI prompt to build an entire [music] AI prompt to build an entire [music] project for me. In reality, I'm project for me. In reality, I'm project for me. In reality, I'm constantly moving between models constantly moving between models constantly moving between models depending on the job. GPT for research, depending on the job. GPT for research, depending on the job. GPT for research, Claude for coding, Gemini for massive Claude for coding, Gemini for massive Claude for coding, Gemini for massive context, [music] context, [music] context, [music] Nano Banana, Midjourney, Flux for Nano Banana, Midjourney, Flux for Nano Banana, Midjourney, Flux for images. And then I've got SeaDance and images. And then I've got SeaDance and images. And then I've got SeaDance and Kling for video. That's why Chat LLM by Kling for video. That's why Chat LLM by Kling for video. That's why Chat LLM by Abacus AI makes sense. It brings day one Abacus AI makes sense. It brings day one Abacus AI makes sense. It brings day one support for the latest GPT, Claude, support for the latest GPT, Claude, support for the latest GPT, Claude, Gemini, Grok, DeepSeek, and more in one Gemini, Grok, DeepSeek, and more in one Gemini, Grok, DeepSeek, and more in one place the moment they drop. Pick any place the moment they drop. Pick any place the moment they drop. Pick any model from the interface or let route model from the interface or let route model from the interface or let route LLM automatically choose the best model LLM automatically choose the best model LLM automatically choose the best model for each prompt. Create professional for each prompt. Create professional for each prompt. Create professional presentations with graphs and charts and presentations with graphs and charts and presentations with graphs and charts and deep research detailed content. Need deep research detailed content. Need deep research detailed content. Need human-sounding copy? Humanize rewrites human-sounding copy? Humanize rewrites human-sounding copy? Humanize rewrites text to defeat AI detectors. Need text to defeat AI detectors. Need text to defeat AI detectors. Need visuals? Pick frontier or open-source

visuals? Pick frontier or open-source visuals? Pick frontier or open-source models. And when you need more than models. And when you need more than models. And when you need more than chat, Abacus AI agent can help build chat, Abacus AI agent can help build chat, Abacus AI agent can help build complex apps and websites, connect complex apps and websites, connect complex apps and websites, connect payments, or run 24/7 agents that keep payments, or run 24/7 agents that keep payments, or run 24/7 agents that keep working through longer tasks. [music] working through longer tasks. [music] working through longer tasks. [music] The best part is app hosting, back-end The best part is app hosting, back-end The best part is app hosting, back-end database, and auth support comes with database, and auth support comes with database, and auth support comes with the subscription. All that starts at the subscription. All that starts at the subscription. All that starts at just $10 a month, way cheaper than just $10 a month, way cheaper than just $10 a month, way cheaper than paying for all those subscriptions paying for all those subscriptions paying for all those subscriptions separately. Check out chat.llm.abacus.ai separately. Check out chat.llm.abacus.ai separately. Check out chat.llm.abacus.ai or click the link below. or click the link below. or click the link below. By the way, how did we get here? Project By the way, how did we get here? Project By the way, how did we get here? Project Volterra, this was 600 bucks. Ah, the Volterra, this was 600 bucks. Ah, the Volterra, this was 600 bucks. Ah, the good old days when computers like this good old days when computers like this good old days when computers like this with 32 gigs of memory were less than with 32 gigs of memory were less than with 32 gigs of memory were less than 1,000 bucks. Snapdragon Oh, if you 1,000 bucks. Snapdragon Oh, if you 1,000 bucks. Snapdragon Oh, if you already know all the history stuff, feel already know all the history stuff, feel already know all the history stuff, feel free to skip around in the chapters. free to skip around in the chapters. free to skip around in the chapters. >> Yeah. >> Yeah. >> Yeah. >> ACX Gen 3, eight cores. This is not even >> ACX Gen 3, eight cores. This is not even >> ACX Gen 3, eight cores. This is not even X Elite yet. Lots of port options. This X Elite yet. Lots of port options. This X Elite yet. Lots of port options. This was a production machine. And I thought was a production machine. And I thought was a production machine. And I thought Project Volterra was a really cool name, Project Volterra was a really cool name, Project Volterra was a really cool name, but they renamed it something silly like but they renamed it something silly like but they renamed it something silly like Windows Dev Kit 2023. But it did its Windows Dev Kit 2023. But it did its Windows Dev Kit 2023. But it did its job. This little box right here is the job. This little box right here is the job. This little box right here is the reason that Visual Studio and .NET went reason that Visual Studio and .NET went reason that Visual Studio and .NET went native on arm. By today's standards native on arm. By today's standards native on arm. By today's standards though, it's slow. I ran it against the though, it's slow. I ran it against the though, it's slow. I ran it against the first X Elite laptop a couple years ago.

first X Elite laptop a couple years ago. first X Elite laptop a couple years ago. Speedometer score was 14 versus 26. My Speedometer score was 14 versus 26. My Speedometer score was 14 versus 26. My Python Mandelbrot test took a minute and Python Mandelbrot test took a minute and Python Mandelbrot test took a minute and a half, but it was only sipping 30 a half, but it was only sipping 30 a half, but it was only sipping 30 watts, and that was it. Then this one, watts, and that was it. Then this one, watts, and that was it. Then this one, the Snapdragon Dev Kit, May 2024, 900 the Snapdragon Dev Kit, May 2024, 900 the Snapdragon Dev Kit, May 2024, 900 bucks. Still pretty decent, I guess. bucks. Still pretty decent, I guess. bucks. Still pretty decent, I guess. First gen X Elite, this was one of the First gen X Elite, this was one of the First gen X Elite, this was one of the top SKUs in the lineup. 32 gigs, I top SKUs in the lineup. 32 gigs, I top SKUs in the lineup. 32 gigs, I ordered it and paid for it as soon as it ordered it and paid for it as soon as it ordered it and paid for it as soon as it was announced, and then Qualcomm was announced, and then Qualcomm was announced, and then Qualcomm canceled it, and they refunded canceled it, and they refunded canceled it, and they refunded everybody, which was nice of them. I got everybody, which was nice of them. I got everybody, which was nice of them. I got to keep the hardware. I put it in my to keep the hardware. I put it in my to keep the hardware. I put it in my seven-way mini PC video anyway. It was seven-way mini PC video anyway. It was seven-way mini PC video anyway. It was the loudest thing on the desk, and it the loudest thing on the desk, and it the loudest thing on the desk, and it pulled 137 watts at the wall on pulled 137 watts at the wall on pulled 137 watts at the wall on Mandelbrot. The M4 Pro Mac mini did the Mandelbrot. The M4 Pro Mac mini did the Mandelbrot. The M4 Pro Mac mini did the same exact thing at 80 watts. But it same exact thing at 80 watts. But it same exact thing at 80 watts. But it beat the M4 Mac mini, and it was the beat the M4 Mac mini, and it was the beat the M4 Mac mini, and it was the only time an arm Windows box has ever only time an arm Windows box has ever only time an arm Windows box has ever beaten the M4 family Mac on this beaten the M4 family Mac on this beaten the M4 family Mac on this channel. And as I said in that video, channel. And as I said in that video, channel. And as I said in that video, many PCs with X Elite chips are going to many PCs with X Elite chips are going to many PCs with X Elite chips are going to hit the market in 2025. You'll see hit the market in 2025. You'll see hit the market in 2025. You'll see I was wrong.

I was wrong. I was wrong. 2026. And now is your chance to actually 2026. And now is your chance to actually 2026. And now is your chance to actually get your hands on one in production. Not get your hands on one in production. Not get your hands on one in production. Not a dev kit, the real thing. This box, 26% a dev kit, the real thing. This box, 26% a dev kit, the real thing. This box, 26% faster on single core and 27% faster on faster on single core and 27% faster on faster on single core and 27% faster on multi-core. This SER 10 here by Beelink, multi-core. This SER 10 here by Beelink, multi-core. This SER 10 here by Beelink, this by the way has the latest AMD chip, this by the way has the latest AMD chip, this by the way has the latest AMD chip, the HX470 the HX470 the HX470 in it. And I also recently did the Asus in it. And I also recently did the Asus in it. And I also recently did the Asus NUC, which has the latest Intel Panther NUC, which has the latest Intel Panther NUC, which has the latest Intel Panther [music] Lake. And the Q and 10 ahead of [music] Lake. And the Q and 10 ahead of [music] Lake. And the Q and 10 ahead of both of them by 20 to 30% both ways. both of them by 20 to 30% both ways. both of them by 20 to 30% both ways. Now, what about that M4 Pro Mac mini? I Now, what about that M4 Pro Mac mini? I Now, what about that M4 Pro Mac mini? I know know know >> [music] >> [music] >> [music] >> M6s are about to come out if they're not >> M6s are about to come out if they're not >> M6s are about to come out if they're not already. Probably already by the time already. Probably already by the time already. Probably already by the time you're watching this. That's a whole you're watching this. That's a whole you're watching this. That's a whole different story and I'm going to have to different story and I'm going to have to different story and I'm going to have to do the comparisons then, but for now, do the comparisons then, but for now, do the comparisons then, but for now, while I'm shooting this video, while I'm shooting this video, while I'm shooting this video, >> [music] >> [music] >> [music] >> the M4 Pro Mac mini is the top of the >> the M4 Pro Mac mini is the top of the >> the M4 Pro Mac mini is the top of the line. The Q and 10 is within a hair of line. The Q and 10 is within a hair of line. The Q and 10 is within a hair of it. And here's the M5 in a MacBook Air. it. And here's the M5 in a MacBook Air. it. And here's the M5 in a MacBook Air. It's not in a mini because there is no It's not in a mini because there is no It's not in a mini because there is no mini with an M5 chip. It's going mini with an M5 chip. It's going mini with an M5 chip. It's going straight to M6. 3597 for single core and straight to M6. 3597 for single core and straight to M6. 3597 for single core and 17,550 for the multi-core score. So, 17,550 for the multi-core score. So, 17,550 for the multi-core score. So, this Q and 10 is actually beating the this Q and 10 is actually beating the this Q and 10 is actually beating the M5. All right. All right. So, Geekbench M5. All right. All right. So, Geekbench M5. All right. All right. So, Geekbench says this little thing is powerful.

says this little thing is powerful. says this little thing is powerful. Fine. Geekbench also said the dev kit Fine. Geekbench also said the dev kit Fine. Geekbench also said the dev kit was powerful and then it lost to a $599 was powerful and then it lost to a $599 was powerful and then it lost to a $599 Mac mini. That's all the stuff we do all Mac mini. That's all the stuff we do all Mac mini. That's all the stuff we do all day. The stuff that matters to day. The stuff that matters to day. The stuff that matters to developers. Let's kick things off with developers. Let's kick things off with developers. Let's kick things off with Speedometer and boom. It basically Speedometer and boom. It basically Speedometer and boom. It basically builds and tears down little to-do apps builds and tears down little to-do apps builds and tears down little to-do apps in every JavaScript framework over and in every JavaScript framework over and in every JavaScript framework over and over again. So, this is a mostly one over again. So, this is a mostly one over again. So, this is a mostly one core operation because JavaScript. This core operation because JavaScript. This core operation because JavaScript. This will tell you how snappy the browser will tell you how snappy the browser will tell you how snappy the browser feels. Woah! 48.6. feels. Woah! 48.6. feels. Woah! 48.6. This is pretty nice. The M4 Mac mini, This is pretty nice. The M4 Mac mini, This is pretty nice. The M4 Mac mini, 47.8. 47.8. 47.8. And up until now, Apple has been leading And up until now, Apple has been leading And up until now, Apple has been leading in single core operations pretty much in single core operations pretty much in single core operations pretty much all along since Apple silicon came out. all along since Apple silicon came out. all along since Apple silicon came out. This is the first time I see a higher This is the first time I see a higher This is the first time I see a higher score. This is pretty amazing. M4 Pro, I score. This is pretty amazing. M4 Pro, I score. This is pretty amazing. M4 Pro, I got 45. We've got Sierra 10, AMD's got 45. We've got Sierra 10, AMD's got 45. We've got Sierra 10, AMD's newest, that's the HX470 chip, and newest, that's the HX470 chip, and newest, that's the HX470 chip, and that's giving us 37. Not bad, but we're that's giving us 37. Not bad, but we're that's giving us 37. Not bad, but we're in a different league now. Next is the in a different league now. Next is the in a different league now. Next is the Web Tooling Benchmark here. It's an Web Tooling Benchmark here. It's an Web Tooling Benchmark here. It's an oldie but a goodie. It's been around on oldie but a goodie. It's been around on oldie but a goodie. It's been around on GitHub for a while. Last commit was 8 GitHub for a while. Last commit was 8 GitHub for a while. Last commit was 8 years ago, but it still works and it years ago, but it still works and it years ago, but it still works and it still gives us an idea how things run in still gives us an idea how things run in still gives us an idea how things run in the JavaScript world and go. Same idea, the JavaScript world and go. Same idea, the JavaScript world and go. Same idea, but instead of running in a web app, it but instead of running in a web app, it but instead of running in a web app, it runs the tools as a CLI. TypeScript, runs the tools as a CLI. TypeScript, runs the tools as a CLI. TypeScript, Babel, Prettier. Basically, Spedometer Babel, Prettier. Basically, Spedometer Babel, Prettier. Basically, Spedometer is like what the user feels and the is like what the user feels and the is like what the user feels and the developer also, but this is really what developer also, but this is really what developer also, but this is really what the developer feels. Babel 6.72, the developer feels. Babel 6.72, the developer feels. Babel 6.72, we've got CoffeeScript 7.76, we've got CoffeeScript 7.76, we've got CoffeeScript 7.76, a bunch of other things that I don't a bunch of other things that I don't a bunch of other things that I don't use. Libab? What the heck is Libab? I've use. Libab? What the heck is Libab? I've use. Libab? What the heck is Libab? I've seen that so many times, but I don't

seen that so many times, but I don't seen that so many times, but I don't know what it is. Do you know what Libab know what it is. Do you know what Libab know what it is. Do you know what Libab is? Prettier 7.25, and the one I'm is? Prettier 7.25, and the one I'm is? Prettier 7.25, and the one I'm waiting for is TypeScript 7.62. Finally, waiting for is TypeScript 7.62. Finally, waiting for is TypeScript 7.62. Finally, geometric mean is 8.15 runs per second. geometric mean is 8.15 runs per second. geometric mean is 8.15 runs per second. These results seem a little bit These results seem a little bit These results seem a little bit disappointing. disappointing. disappointing. Gotcha. You're surprised? This right Gotcha. You're surprised? This right Gotcha. You're surprised? This right here is exactly what happens when you here is exactly what happens when you here is exactly what happens when you run an X64 program on an ARM-based chip run an X64 program on an ARM-based chip run an X64 program on an ARM-based chip in Windows. It runs. It's running in Windows. It runs. It's running in Windows. It runs. It's running through prism translation, which is that through prism translation, which is that through prism translation, which is that layer. It's similar to how Rosetta was layer. It's similar to how Rosetta was layer. It's similar to how Rosetta was on Apple silicon. It translates X86 on Apple silicon. It translates X86 on Apple silicon. It translates X86 programs on the fly, but the performance programs on the fly, but the performance programs on the fly, but the performance degrades. And that goes for Node, my degrades. And that goes for Node, my degrades. And that goes for Node, my friends, and I did not install the friends, and I did not install the friends, and I did not install the correct version of Node. You can find correct version of Node. You can find correct version of Node. You can find out which version you're running by out which version you're running by out which version you're running by doing this, or you can just ask Claude doing this, or you can just ask Claude doing this, or you can just ask Claude code, but this is the old school way of code, but this is the old school way of code, but this is the old school way of doing it, and it says X64 right here. doing it, and it says X64 right here. doing it, and it says X64 right here. So, let's see what happens when we So, let's see what happens when we So, let's see what happens when we install the right version of Node. I got install the right version of Node. I got install the right version of Node. I got the ARM64 version of Node installed in a the ARM64 version of Node installed in a the ARM64 version of Node installed in a directory, and there it is. V24, and the directory, and there it is. V24, and the directory, and there it is. V24, and the process is an ARM64 process. So, we're process is an ARM64 process. So, we're process is an ARM64 process. So, we're good to go on that. Huh, I'm running it, good to go on that. Huh, I'm running it, good to go on that. Huh, I'm running it, but it's a little bit faster.

but it's a little bit faster. but it's a little bit faster. >> It's not faster enough. >> It's not faster enough. >> It's not faster enough. >> 11.8 runs per second. The Sierra 10 does >> 11.8 runs per second. The Sierra 10 does >> 11.8 runs per second. The Sierra 10 does 28. The old dev kit, the one they 28. The old dev kit, the one they 28. The old dev kit, the one they canceled, did 22. This chip just beat canceled, did 22. This chip just beat canceled, did 22. This chip just beat the M4 on Speedometer. And that was the M4 on Speedometer. And that was the M4 on Speedometer. And that was giving me the performance of 1/3 of the giving me the performance of 1/3 of the giving me the performance of 1/3 of the AMD box running the same exact AMD box running the same exact AMD box running the same exact JavaScript. JavaScript. JavaScript. >> It's just not faster enough. >> It's just not faster enough. >> It's just not faster enough. >> So, here's what's going on as far as I >> So, here's what's going on as far as I >> So, here's what's going on as far as I can tell. Speedometer runs inside can tell. Speedometer runs inside can tell. Speedometer runs inside Chrome. This runs inside Node. It's the Chrome. This runs inside Node. It's the Chrome. This runs inside Node. It's the V8 engine, but has a different wrapper. V8 engine, but has a different wrapper. V8 engine, but has a different wrapper. First time I ran it, we went from eight First time I ran it, we went from eight First time I ran it, we went from eight on X64 version of Node to 11 on the arm on X64 version of Node to 11 on the arm on X64 version of Node to 11 on the arm version of Node, which is better, but version of Node, which is better, but version of Node, which is better, but it's still not where it should be. But it's still not where it should be. But it's still not where it should be. But luckily, this test also is available as luckily, this test also is available as luckily, this test also is available as an HTML page. Here it is, which opens up an HTML page. Here it is, which opens up an HTML page. Here it is, which opens up the test in Chrome. Let's run this. the test in Chrome. Let's run this. the test in Chrome. Let's run this. Boom. Boom. Boom. There we go. Yeah, now we're cooking There we go. Yeah, now we're cooking There we go. Yeah, now we're cooking with butter. with butter. with butter. Ah, TypeScript 63. Geometric mean 49.89. Ah, TypeScript 63. Geometric mean 49.89. Ah, TypeScript 63. Geometric mean 49.89. That's what I'm talking about. Wow, That's what I'm talking about. Wow, That's what I'm talking about. Wow, okay. This box beats everything else by okay. This box beats everything else by okay. This box beats everything else by quite a lot. But just so you know, we're quite a lot. But just so you know, we're quite a lot. But just so you know, we're comparing the Chrome run comparing the Chrome run comparing the Chrome run to the Node CLI run. Not exactly to the Node CLI run. Not exactly to the Node CLI run. Not exactly apples-to-apples, but it does look apples-to-apples, but it does look apples-to-apples, but it does look impressive. Now, JavaScript is one impressive. Now, JavaScript is one impressive. Now, JavaScript is one thing, but what about compiled code?

thing, but what about compiled code? thing, but what about compiled code? Well, 100,000 namespaces, that's what I Well, 100,000 namespaces, that's what I Well, 100,000 namespaces, that's what I got over here, and each one with real got over here, and each one with real got over here, and each one with real code inside, so the compiler can't code inside, so the compiler can't code inside, so the compiler can't really skip it. This is the test that I really skip it. This is the test that I really skip it. This is the test that I created specifically as kind of a large created specifically as kind of a large created specifically as kind of a large compilation in .NET. So, let's run this. compilation in .NET. So, let's run this. compilation in .NET. So, let's run this. Now, 2 years ago, the dev kit won this Now, 2 years ago, the dev kit won this Now, 2 years ago, the dev kit won this test, 86 seconds to M4's 106 seconds. test, 86 seconds to M4's 106 seconds. test, 86 seconds to M4's 106 seconds. What's the new one going to get? So, What's the new one going to get? So, What's the new one going to get? So, what I want to get a sense of how quiet what I want to get a sense of how quiet what I want to get a sense of how quiet it is and how much power it draws. I it is and how much power it draws. I it is and how much power it draws. I guess we do need that big power brick. guess we do need that big power brick. guess we do need that big power brick. 109 watts is what I've seen it hit, but 109 watts is what I've seen it hit, but 109 watts is what I've seen it hit, but this machine is very quiet. It's warm, this machine is very quiet. It's warm, this machine is very quiet. It's warm, sure, but it hasn't made any noise. Very sure, but it hasn't made any noise. Very sure, but it hasn't made any noise. Very nice little companion to be on a desk nice little companion to be on a desk nice little companion to be on a desk without making noise. I like that. And without making noise. I like that. And without making noise. I like that. And the build succeeded in 77 seconds. the build succeeded in 77 seconds. the build succeeded in 77 seconds. Pretty good. However, we did not beat Pretty good. However, we did not beat Pretty good. However, we did not beat the M4 Pro Mini. That one got 66 the M4 Pro Mini. That one got 66 the M4 Pro Mini. That one got 66 seconds. But this thing did come in seconds. But this thing did come in seconds. But this thing did come in second. And this is a multi-core second. And this is a multi-core second. And this is a multi-core compilation. Now, besides compilations, compilation. Now, besides compilations, compilation. Now, besides compilations, I have an interpreted type of test here, I have an interpreted type of test here, I have an interpreted type of test here, which is called the Mandelbrot test. which is called the Mandelbrot test. which is called the Mandelbrot test. It's a Mandelbrot algorithm. I got this It's a Mandelbrot algorithm. I got this It's a Mandelbrot algorithm. I got this code from Benchmark game. Boom. I'm code from Benchmark game. Boom. I'm code from Benchmark game. Boom. I'm using measure command here in PowerShell using measure command here in PowerShell using measure command here in PowerShell to get the timing of that run. Now, this to get the timing of that run. Now, this to get the timing of that run. Now, this basically is just doing the Mandelbrot basically is just doing the Mandelbrot basically is just doing the Mandelbrot algorithm, which is drawing fractals in algorithm, which is drawing fractals in algorithm, which is drawing fractals in Python code. And this really hits the Python code. And this really hits the Python code. And this really hits the CPU hard. Look at all those CPU cores.

CPU hard. Look at all those CPU cores. CPU hard. Look at all those CPU cores. Utilization is at 100%. Oh, Utilization is at 100%. Oh, Utilization is at 100%. Oh, I'm starting to hear it a little bit. I'm starting to hear it a little bit. I'm starting to hear it a little bit. The fans have kicked in, but it's still The fans have kicked in, but it's still The fans have kicked in, but it's still very quiet. And it's done. 50.43 very quiet. And it's done. 50.43 very quiet. And it's done. 50.43 seconds. seconds. seconds. Uh Uh Uh there. there. there. That's a bit lower than I was expecting. That's a bit lower than I was expecting. That's a bit lower than I was expecting. Again, we have the same issue. What Again, we have the same issue. What Again, we have the same issue. What version of Python do we have here? version of Python do we have here? version of Python do we have here? Yep. This is the X64 version of Python. Yep. This is the X64 version of Python. Yep. This is the X64 version of Python. It'll bite you. It'll bite you. It'll bite you. Be careful. But hey, now at least you Be careful. But hey, now at least you Be careful. But hey, now at least you know what happens. And think about it know what happens. And think about it know what happens. And think about it this way, you got a two-for-one this way, you got a two-for-one this way, you got a two-for-one comparison for free. And let's go. Boom. comparison for free. And let's go. Boom. comparison for free. And let's go. Boom. See, I'm running now the new version of See, I'm running now the new version of See, I'm running now the new version of Python, which I put in here. That's the Python, which I put in here. That's the Python, which I put in here. That's the arm executable. arm executable. arm executable. >> [laughter] >> [laughter] >> [laughter] >> Is that already? 21.76 >> Is that already? 21.76 >> Is that already? 21.76 seconds. Now, this is pretty crazy seconds. Now, this is pretty crazy seconds. Now, this is pretty crazy because it won. It actually beat all the because it won. It actually beat all the because it won. It actually beat all the other machines. Wow. We're pulling about other machines. Wow. We're pulling about other machines. Wow. We're pulling about 120 to 125 watts. 21.8 on the QN 10. The 120 to 125 watts. 21.8 on the QN 10. The 120 to 125 watts. 21.8 on the QN 10. The M4 Pro just a little bit slower at 23.2.

M4 Pro just a little bit slower at 23.2. M4 Pro just a little bit slower at 23.2. Both the XLE dev kit and the Ser 10 got Both the XLE dev kit and the Ser 10 got Both the XLE dev kit and the Ser 10 got 28.9, and the M4 Mac Mini 31.4. And I 28.9, and the M4 Mac Mini 31.4. And I 28.9, and the M4 Mac Mini 31.4. And I threw in the uh Python emulation threw in the uh Python emulation threw in the uh Python emulation on this box. 50.4 seconds. Just so that on this box. 50.4 seconds. Just so that on this box. 50.4 seconds. Just so that you're aware and you remember. But wow, you're aware and you remember. But wow, you're aware and you remember. But wow, this thing is actually fast. You know this thing is actually fast. You know this thing is actually fast. You know what we haven't done yet is we haven't what we haven't done yet is we haven't what we haven't done yet is we haven't done any AI, which we're going to do done any AI, which we're going to do done any AI, which we're going to do right now. This box is supposed to have right now. This box is supposed to have right now. This box is supposed to have 80 tops, and that's the NPU. So, I went 80 tops, and that's the NPU. So, I went 80 tops, and that's the NPU. So, I went and got llama.cpp with the Hexagon and got llama.cpp with the Hexagon and got llama.cpp with the Hexagon backend. Hexagon is the NPU, but it backend. Hexagon is the NPU, but it backend. Hexagon is the NPU, but it never got past the driver. I just could never got past the driver. I just could never got past the driver. I just could not get it working. Uh it's probably not get it working. Uh it's probably not get it working. Uh it's probably something I did. Yeah, most likely. Let something I did. Yeah, most likely. Let something I did. Yeah, most likely. Let me know in the comments down below if me know in the comments down below if me know in the comments down below if you got the X2 or the X1 Elite Hexagon you got the X2 or the X1 Elite Hexagon you got the X2 or the X1 Elite Hexagon NPU working with llama.cpp and I'd be NPU working with llama.cpp and I'd be NPU working with llama.cpp and I'd be curious to hear from you. So, I moved on curious to hear from you. So, I moved on curious to hear from you. So, I moved on to the other things that matter. Next to the other things that matter. Next to the other things that matter. Next was the GPU. This machine has unified was the GPU. This machine has unified was the GPU. This machine has unified memory, 32 gigs of it. So, you'd think memory, 32 gigs of it. So, you'd think memory, 32 gigs of it. So, you'd think that a decently sized model would just that a decently sized model would just that a decently sized model would just load. Well, it doesn't. Through OpenCL, load. Well, it doesn't. Through OpenCL, load. Well, it doesn't. Through OpenCL, the GPU is only allowed to use about 15 the GPU is only allowed to use about 15 the GPU is only allowed to use about 15 gigs of that memory. The other 17 gigs gigs of that memory. The other 17 gigs gigs of that memory. The other 17 gigs are sitting right there and the GPU are sitting right there and the GPU are sitting right there and the GPU can't touch them. They're reserved. So, can't touch them. They're reserved. So, can't touch them. They're reserved. So, anything big ends up running on the CPU anything big ends up running on the CPU anything big ends up running on the CPU instead. And that actually turned out to instead. And that actually turned out to instead. And that actually turned out to be fine in this case. I wanted to start be fine in this case. I wanted to start be fine in this case. I wanted to start out simple. So, on llama 3B, the CPU out simple. So, on llama 3B, the CPU out simple. So, on llama 3B, the CPU does 41 tokens a second and the GPU does does 41 tokens a second and the GPU does does 41 tokens a second and the GPU does 31. GPT-OSS-20B, 35 on the CPU and 33 on 31. GPT-OSS-20B, 35 on the CPU and 33 on 31. GPT-OSS-20B, 35 on the CPU and 33 on the GPU. Usually, it's flipped around, the GPU. Usually, it's flipped around, the GPU. Usually, it's flipped around, but here it's not. In fact, every single but here it's not. In fact, every single but here it's not. In fact, every single model I tried, the CPU was faster than model I tried, the CPU was faster than model I tried, the CPU was faster than the GPU. I didn't expect that. For the the GPU. I didn't expect that. For the the GPU. I didn't expect that. For the uninitiated, let me give you a quick

uninitiated, let me give you a quick uninitiated, let me give you a quick feel of what those numbers mean. Around feel of what those numbers mean. Around feel of what those numbers mean. Around five tokens a second is about as fast as five tokens a second is about as fast as five tokens a second is about as fast as you can read. So, you and the model you can read. So, you and the model you can read. So, you and the model finish the paragraph together. I read finish the paragraph together. I read finish the paragraph together. I read much slower myself, so much slower myself, so much slower myself, so uh it's going to take me a little bit uh it's going to take me a little bit uh it's going to take me a little bit longer. That's why I'm okay with slower longer. That's why I'm okay with slower longer. That's why I'm okay with slower speeds, I guess. But for those geniuses speeds, I guess. But for those geniuses speeds, I guess. But for those geniuses out there, at 40 tokens a second, the out there, at 40 tokens a second, the out there, at 40 tokens a second, the words are coming out faster than you can words are coming out faster than you can words are coming out faster than you can skim them. Unless you are said genius, skim them. Unless you are said genius, skim them. Unless you are said genius, then you can keep up. And at 100 tokens then you can keep up. And at 100 tokens then you can keep up. And at 100 tokens per second, the whole answer is just per second, the whole answer is just per second, the whole answer is just done before you even look up. So, keep done before you even look up. So, keep done before you even look up. So, keep that in mind when you look at the next that in mind when you look at the next that in mind when you look at the next few charts. Now, let's bring in the few charts. Now, let's bring in the few charts. Now, let's bring in the other machines on the desk. I ran Llama other machines on the desk. I ran Llama other machines on the desk. I ran Llama 3.23B, the same model I used in the 3.23B, the same model I used in the 3.23B, the same model I used in the seven-way video from last year or 2 seven-way video from last year or 2 seven-way video from last year or 2 years ago. years ago. years ago. Anyway, time is flying when you're Anyway, time is flying when you're Anyway, time is flying when you're having so much fun doing this stuff. The having so much fun doing this stuff. The having so much fun doing this stuff. The Dev Kit did about 36 tokens a second. Dev Kit did about 36 tokens a second. Dev Kit did about 36 tokens a second. This box does 41. The M4 Mac Mini does This box does 41. The M4 Mac Mini does This box does 41. The M4 Mac Mini does 46, and the M4 Pro does 101. Now, we're 46, and the M4 Pro does 101. Now, we're 46, and the M4 Pro does 101. Now, we're going to get the M5 Pros, which are going to get the M5 Pros, which are going to get the M5 Pros, which are going to be going to be going to be even better. But, the QN10 has pretty even better. But, the QN10 has pretty even better. But, the QN10 has pretty decent performance for its price. And decent performance for its price. And decent performance for its price. And also, the Macs are running MLX, which is also, the Macs are running MLX, which is also, the Macs are running MLX, which is Apple's framework. It's their best use Apple's framework. It's their best use Apple's framework. It's their best use case scenario. Why does the M4 Pro pull case scenario. Why does the M4 Pro pull case scenario. Why does the M4 Pro pull so far ahead? Well, that's not just so far ahead? Well, that's not just so far ahead? Well, that's not just software. That's the memory bandwidth.

software. That's the memory bandwidth. software. That's the memory bandwidth. And the M4 Pro has a much higher memory And the M4 Pro has a much higher memory And the M4 Pro has a much higher memory bandwidth than the X2 Elite. Now, big, bandwidth than the X2 Elite. Now, big, bandwidth than the X2 Elite. Now, big, dense models dense models dense models are really rough on this machine. 27B are really rough on this machine. 27B are really rough on this machine. 27B models run at about five or six tokens a models run at about five or six tokens a models run at about five or six tokens a second, which is basically your reading second, which is basically your reading second, which is basically your reading speed. With dense models, every single speed. With dense models, every single speed. With dense models, every single parameter wakes up for every word parameter wakes up for every word parameter wakes up for every word generated. So, that's why it's really generated. So, that's why it's really generated. So, that's why it's really heavy. Mixture of experts models are a heavy. Mixture of experts models are a heavy. Mixture of experts models are a different story. They only wake up a different story. They only wake up a different story. They only wake up a subset of parameters for every word. So, subset of parameters for every word. So, subset of parameters for every word. So, it feels like you're running a smaller it feels like you're running a smaller it feels like you're running a smaller model, even though they're big, and it's model, even though they're big, and it's model, even though they're big, and it's faster. GPT-OSS-20B is an example of faster. GPT-OSS-20B is an example of faster. GPT-OSS-20B is an example of mixture of experts model, MoE. It runs mixture of experts model, MoE. It runs mixture of experts model, MoE. It runs at 35 tokens a second, and Qwen 3.5 35B at 35 tokens a second, and Qwen 3.5 35B at 35 tokens a second, and Qwen 3.5 35B runs at 27. That's one heck of a CPU. runs at 27. That's one heck of a CPU. runs at 27. That's one heck of a CPU. That Qwen model is 22 GB. It fits on That Qwen model is 22 GB. It fits on That Qwen model is 22 GB. It fits on this box, and it does not fit on a 16 this box, and it does not fit on a 16 this box, and it does not fit on a 16 gig Mac Mini. If you're wondering why gig Mac Mini. If you're wondering why gig Mac Mini. If you're wondering why you'd need 32 gigs, well, that's a good you'd need 32 gigs, well, that's a good you'd need 32 gigs, well, that's a good reason. Now, two things to know in a reason. Now, two things to know in a reason. Now, two things to know in a more realistic scenario. If you paste in more realistic scenario. If you paste in more realistic scenario. If you paste in a big file, if you have a large context, a big file, if you have a large context, a big file, if you have a large context, around 16,000 tokens, you're going to around 16,000 tokens, you're going to around 16,000 tokens, you're going to wait 53 seconds before the first word wait 53 seconds before the first word wait 53 seconds before the first word shows up. And if your buddy wants to use shows up. And if your buddy wants to use shows up. And if your buddy wants to use the same computer, the same model at the the same computer, the same model at the the same computer, the same model at the same time as you, same time as you, same time as you, and starts using it, and starts using it, and starts using it, everybody drops down to about seven everybody drops down to about seven everybody drops down to about seven tokens a second. This is basically a tokens a second. This is basically a tokens a second. This is basically a machine for one person sitting at a machine for one person sitting at a machine for one person sitting at a desk, and for that person, it's a good desk, and for that person, it's a good desk, and for that person, it's a good one. So, is this the Mac mini killer?

one. So, is this the Mac mini killer? one. So, is this the Mac mini killer? Well, if you're a Windows developer, Well, if you're a Windows developer, Well, if you're a Windows developer, then yes. It kept the promise that the then yes. It kept the promise that the then yes. It kept the promise that the dev kit made, and I've been waiting for dev kit made, and I've been waiting for dev kit made, and I've been waiting for 2 years to have the Snapdragon chip 2 years to have the Snapdragon chip 2 years to have the Snapdragon chip inside one of these mini PCs. Now, the inside one of these mini PCs. Now, the inside one of these mini PCs. Now, the base model Mac mini is still a cheaper base model Mac mini is still a cheaper base model Mac mini is still a cheaper option. It's still faster on the small option. It's still faster on the small option. It's still faster on the small models with MLX. This box costs a bit models with MLX. This box costs a bit models with MLX. This box costs a bit more. So, for AI purposes, more. So, for AI purposes, more. So, for AI purposes, I would skip this box. It does have that I would skip this box. It does have that I would skip this box. It does have that 18 cores and 32 gigs, though. So, if 18 cores and 32 gigs, though. So, if 18 cores and 32 gigs, though. So, if you're doing Windows development, if you're doing Windows development, if you're doing Windows development, if you're doing other development other you're doing other development other you're doing other development other than AI, this is a really good choice. than AI, this is a really good choice. than AI, this is a really good choice. Just keep in mind arm. Make sure your Just keep in mind arm. Make sure your Just keep in mind arm. Make sure your software works on arm. Now, if you want software works on arm. Now, if you want software works on arm. Now, if you want to see a seven-way mini PC showdown with to see a seven-way mini PC showdown with to see a seven-way mini PC showdown with that dev kit, watch this video right that dev kit, watch this video right that dev kit, watch this video right over here. Thanks for watching, and I'll over here. Thanks for watching, and I'll over here. Thanks for watching, and I'll see [music] you next time.