@IoanBiTza Snapdragon X Elite also use the ARM architecture, but they can't even release a fanless model, because it's not enough efficient.. loosing half of the performance without a cooler..
The things I enjoy most on the MBP are the total lack of fan noise and the fact that I never have to tweak performance settings. It just runs buttery smooth plugged in or on battery. And the battery life is actually great!
Honestly, calling the 14900HX a "brand new processor" in 2025 is borderline disingenuous. It's basically a rebranded 13th gen CPU which is more than 2 years old. Also, the RTX 4090 launched in 2022, and the RTX 50-series was recently announced. The newly announced AMD AI Max+ 395 (yes, awful name) chip would be a much more interesting comparison. It is a "all in one" chip more similar to the Apple Silicon ones. They are expected to be decent value and power efficient too!
@@Berecutecu that was the whole point I made, the only reason you might even consider Intel now is if you NEED TB5. Before Strix Halo, it could brute force multi-thread stuff slightly better than AMD because it had more cores... that's no longer the case.
There's something screwy about that Razor. I ran Llama3.2 on my ancient RTX2080 laptop and I got 123 tokens/s. No way can a 4090 only do a 1-2% increase over a card that's two generations older and a lot less powerful. And that multicore score on Geekbench is abysmal.
The reason that both the M4 Pro and M4 Max were similar in performance may be due to the Neural Engine being identical in both chips. If properly managed, most of the network, if not all, should run on the ANE, thus providing similar results. It would be very interesting to see where the model runs during inference with Asitop (which shows CPU, GPU and ANE usage). Great comparison!
Just to be clear, they weren’t similar in performance. Using MLX-based runtimes, the M4 Max achieved 172 tokens/second to 103 tokens/second for the M4 Pro. The difference is most likely due to memory bandwidth.
Last time I checked, LLMs were not able to be run on the ANE. Maybe that changed by now, but there was something about Apple not providing 3rd party devs the necessary APIs to use it in full.
@ There is no limitation preventing LLMs from running on the ANE, it comes down to the architecture implementation and quantization. What is interesting though is that MLX, currently does not support the ANE, at least according to python MLX package on PYPI. ANE can be fully utilized though using CoreML, even for 3rd parties (e.g. coremltools on python). On my models, I can see significant speed boosts when running on ANE with CoreML.
I’ve always been a Windows user, rocking an RTX 1070m 3080m, 3090. But when the time came to choose between the Lenovo Legion Pro 7 (AMD Ryzen 9 7940HX or Intel Core i9-14900HX) for $3,400, or the MacBook Pro 16" M4 Max (48GB RAM, 1TB SSD) for $3,229 at my college bookstore, I went with the MacBook. At first, I was skeptical I'd always thought Apple products were overpriced and overrated, especially based on things like AirPods Max and every flashy new iPhone. But oh boy, was I wrong. This MacBook is a powerhouse. It handles my 4K and even 8K drone footage in Final Cut Pro effortlessly, staying almost ice-cold and performing incredibly fast. Whether it’s 3D modeling or programming in IntelliJ, this machine is a beast. As a gamer with over 300 games on Steam, I was initially concerned about the limited macOS library with only half of all my games available, but I’ve been pleasantly surprised. Here are the games I’ve run on 4K ultra settings so far: - **Metro Exodus**: 100-120 FPS - **Dying Light**: 120 FPS (200+ FPS with VSync off) - **Hogwarts Legacy** (via Crossover): 90-100 FPS - **Elite Dangerous** (via Parallels Desktop): 70 FPS - **No Man's Sky**: 120 FPS - **Baldur’s Gate 3**: 100-120 FPS - **The Witcher 3: Wild Hunt** (via Crossover): 90-100 FPS And there’s more! What’s impressed me the most is how quiet the MacBook stays. Even under heavy loads, the fans barely spin up, and at times, they don’t spin at all. The laptop remains completely silent with no FPS drops. Whatever Apple is doing with their hardware and software integration, they’re absolutely on the right track-and I’m confident it’s only going to get better from here. I have zero regrets about making the switch.
@@AZisk could you please explain in one of your future videos about FP4 vs FP16, because if I understand correct NVIDIA Project DIGITS is around 1000 TOPS vs RTX 5090 which is 3.1 TOPS so the 5090 is around 3 Petaflobs? (BTW I'm not native English speaker I'm sorry if any mistake)
Hi Alex, did you try changing the performance modes in the razer synapse software? That normally controls the tdp for both the GPU and CPU on Razer laptops. I would also recommend changing it to discrete only mode for the GPU as it would run the tests on the 4090
I have MBP and an old 4 year old Ryzen linux laptop, the Windows VM on Linux boots my dev env app in 5 seconds, the MBP with ARM Windows boots my env in almost 58 seconds and that's using paid Parallels...I can't stand my mbp from a dev perspective. The way keyboard "shortcuts" are implemented sucks as it's not always supported in all apps.
I've bought two 15 inch Razer laptops, 2020 and 2021 Advanced models, both had suboptimal cooling solutions for the compnents they pack and constantly overheated. The other bug bear was needing to run their Razer Synapse software in the background to be able to control keyboard lighting, fan speeds and performance modes. This coupled with their poor build & design quality I honestly can't say I would recommend them.
@@migueljardim8177 I owned a 2022 model and the edges of this laptop was way to sharp and the keyboard is one of the worst i ever used. Now its sold and replaced with a ThinkPad.
@@migueljardim8177 I've had to open both to resolve multiple issues and I was not impressed at all! And don't get me started on trying to find replacement components like fans, keyboards or touchpads when they die!
A couple of things to note for ML benchmark: first I don’t think LM studio works good for nvidia chips. Usually we use vllm or sglang to serve the model which is way faster than LM studio. And the benchmarks better to be done with bigger model to exhaust power of 4090 compared to just 3B model considering it’s already 4-bit quantised. Second thing is about inference throughput. Apple mlx has never done any good batch LLM inference, meaning if you have multiple queries, it’s going to be slow to process. I can easily reach 2k tokens/second for a mobile 4090 for a 3B model with batch inference.
Yeah - I'm serving myself a local LLM on my homeserver, and while setting things up I learned that llama.cpp is really not that fast or efficient when fully loading models up into the VRAM. Things like vllm and exllamav2 with tabby or aphrodite-engine were recommended instead.
The goal is not to provide a balanced analysis of what the machines can potentially do if you use them properly. It all about shitting on PCs to make the Mac look as the superior choice no matter what. This is to lower the cognitive dissonance associated with buying very expensive hardware that isn't that much better in the real world. Hence the sensationalist titles and whatnot. This guy keeps showing up in my recommended, yet every time his approach makes no sense appart from just shitting on PCs. To be clear I own Mac/iPhone/aWatch. There is some good thing about it but the value proposition is out of whack...
@@clementcollier8432 but both fall under the category of very expensive hardware. It's hardly just about justifying the price. Gonna be honest: I don't like the MacOS. I don't even have legitimate reasons, I just dislike using it. But I do think that as things stand now, Macbooks are better for most people at least as general purpose laptops, but also for some very specific work tasks(video/foto editing, coding, audio work etc). If you need the best GPU on the market and CUDA accelerated applications, obviously that is only present on the Windows side. But looking at the device as a whole, it just looks like a horrible proposition to me - it's loud, it uses a ton of power as it's inefficient, it drops in performance and by a significant margin when used on battery (which btw was very generous of him to use it plugged in basically the whole time), consequently have poor battery life. Granted, it's a poor hardware choice and I would like to see a comparable AMD based device with some of their upcoming offerings, but these Intel machines are pretty bad. "There is some good thing about it but the value proposition is out of whack"....hm, not sure this criticism can be directed towards the Mac/Macbook lineup. They are pretty on par with the Windows offerings in terms of value. Like sure they ain't cheap, but on the other side, the premium Windows devices aren't cheap either. As soon as you hit any sort of overall equivalency in terms of build quality, battery, screen quality, and somewhat comparable performance, you're basically looking at the same price range. This can be more directed towards the smaller devices, namely iPhones. And there's also the new, very cheap but capable Mac Mini now. It's a lot of PC for the price, to the extent that if you don't intend to do any sort of gaming, it excels at the price point compared to the competition which doesn't have anything to offer in the price range in that form factor at that performance level. And because of my aforementioned dislike of MacOS, I've been following the market closely for the past year or so to see if something will pop up to satisfy my desire for a new non-MacOS laptop but so far, everything that showed any promise turned out to cost the same as a comparable Macbook Pro, but also turned out to have some sort of flaw that would just annoy me. In some cases they even cost more. I'm not in a rush and will wait to give AMD a chance to prove me wrong, but at this point the internals aren't even that much of an issue, but the whole package. The thing is, I may dislike their OS, but if I get a Macbook, I at least know what I am getting. It's a safe bet. Windows laptops are so tricky to get a good read on, since the market offerings are way too granular and the product lines are very messy for all of the manufacturers.
Now would not be a good time to get a Blade laptop since very soon they'll be refreshed with the Nvidia 50 series GPU's and new AMD and Intel CPU's. I'd wait personally, especially for the AMD Strix Halo chips.
@@migueljardim8177 When new game laptop models come out I buy last years model marked 80% off. FWIW I don’t value AI frames as highly as rendered frames. The 5 series performance is numbers not experience. Why pay 10X as much so the card can hallucinate it is preforming???
I mean Razer has got it's own proprietary software called Razer synapse or something that lets you set it to quiet balance, performance mode etc. so.. maybe have a look there, that windows performance setting is just useless for gaming laptops
Small detail regarding ML test: compute increases approximately quadratically with number of tokens. So if the output on the two machines is not identical in length, there's a small bias. Also when you prompt again in the same context, more compute is needed.
In other words, when you output e.g. 900 tokens in one test and 1000 tokens in another test, there should be slightly more tokens per second in the former case.
I really would love to see a comparison between the M4 Max and the 40x0 mobile for ML/Pytorch training. I am pretty sure you did a benchmark in one of your other videos, and I found it super useful. I know inference could be a proxy.. but imo.. training can behave differently As always Alex, thanks for the video - great work as always!
Im noticing a big divide between next gen machines and everything else. Fast RAM and unified memory I think is making a staggering difference. Plus as you mentioned storage speed and bad cpu's. Mac is amazing. I cant wait for the AMD AI Max+ 395, that might be the closest Mac competitor we'll see.
16 Zen 5 cores (32 threads) and 128 gb of RAM will be INSANE. It will even destroy the 1 thing Intel was decent at (multi-core because they shoved so many cores in their silicon). Unless you need TB5, the choice is clear.
I did not expect those results. Wow. I'm in the market for a new laptop for data analytics work. Going back to Mac wasn't even on my radar. Holy cow I have to rethink this now. GREAT VIDEO. Thanks for complicating my choices. Haha.
For the drive, on the PC does it show up? Check the format drive application and if it doesn’t show up there try using ‘diskpart’ from the command line if you didn’t already. Most APFS formatting doesn’t show up on Windows unless you do a command line level diskpart
I don't know if you have ever done this before, but I would love to see some comparisons running Linux. For some of these tests, I really wonder if Windows is a problem.
@@AZisk hopefully at least some kind of competition. As much as I dislike MacOS this hardware is more than making up for OS. And so far its not even close.
Any chance you throw a Linux distro on the Razor and re-run the tests? Obviously the disk performance will be better for things like that compilation test (NTFS is awful with lots of small files), but I’d be curious to see the difference in the AI benchmarks. Oh, and if you want a real-world compilation test for .Net, why not the .Net framework itself?
Regarding the terrible write speeds on your thunderbolt 5 enclosure in windows you have to go to device manager -> disk drives -> right click on the drive -> properties -> go to the policies tab and enable high performance and enable write caching on the device. Now it should perform as expected.
This is a channel dedicated to developers, so it's not weird at all. Half the video was about compiling code, and the other was about LLM testing. It has covered the bases there. Several channels focus on photography and videography that show the MacBooks easily beating Intel, so watch them if that is your thing. Otherwsie, what pro workload should be tested?
I think that in selecting 4 bit small models so they fit on the 4090 VRAM you are teeing up the MacBooks to run on the NPU. The reason I think this is that the Pro and Max chips performed the same. ~40 tops is very capable for any 4bit inference that will fit in memory. For the 4090 to outperform the M4 you would need to do some 16bit and or learning. The other big advantage of the M4 is that it automatically sends instructions to the best core and they all call the same unified memory. The i9 is doing nothing but management of the 4090. For 4 bit inference the i9 might be about as good as the 4090.
Well , you weren't testing it for Gaming which is what the Razer Blade is all about. But still pretty startling. I think both the energy/heat envelope plus the faster (and unified) memory on the Mac is the big differentiator. Can't be the raw CPUs and GPUs, can it?
I've heard about the price of the razer... what about the MacBook pro? In any case, yes, the M4 Max is a beast and clearly shows the benefits of designing the CPU/hardware and operating system in house.
And we're not even talking about the build quality, speakers etc and ofc windows itself. This is just embarrassing. It's really only "good" for gaming and even at that it's bad, lots of noise and awful performance for the price. Especially on battery, might as well get a desktop otherwise.
I own a Blade 16 from last year and I am not offended, Alex. 😆 However...it is well known that Razer generally builds expensive machines mostly intended for gaming, not for computing LLM and string builds. As such, for a dev like you the new M4 offerings by Apple are your nirvana of course. ✌
14:41 yes and no, The unified memory is faster than non unified , BUT in this case , the razor 18 (I think) has sodim up gradable memory, that is slower than soldered down memory.
Windows devices are being held back by their OS and windows defender, before running any task it just starts first then lets the app to run which means extremely slow speed compared to other OS. If you run this device on Linux results will be a lot different
I can't wait for Strix Halo chips to hit the market, I want to see how the AI Max 395+ (naming scheme sucks) compares to the M4. Both are SoC's that have integrated RAM and all that, so it should be a more apples to apples comparison.
I have had some issues with consistency from a benchmark called silverbench but it would be cool to add it to your testing because it can be run on computers on demo via any web browser. That or any other benchmark in the browser.
even if you go to settings and turn on best performance, this is still going to be based off of the power plan settings set in the ctrl panel, go to cmd prompt and turn on ultimate performance power plan, and please post the scores from there, I'd like to see those scores as well.
The fan noise on that Razer laptop alone is a deal-breaker for me. My Mac programming needs are simple (C++, web) so I use a 15" M2 MacBook Air with 24GB that I purchased on launch day in June 2023. For me the best features are the 15" screen, the travel size and weight, the ergonomics - large trackpad and palm rests, and ... NO FANS. I think that last one puts the MacBook Air in a class by itself. Are there any Windows laptops without fans? If there were, that would be an interesting comparison.
@Harshcodes2 true, windows runs a minority is internet servers. But Windows does run the majority of corporate servers. I recognize Macs are good machines for end users in some scenarios.
Please use larger models with 14B such as qwen2.5 14B. These small models are too small to get a meaningful result also another important metric is prompt ingestion t/s. how fast it reads the prompt. A 3 words prompt is too small to get a meaningful response. Basically please do more meaningful, real world LLM tests.
11:35 Mac has 128 GB RAM vs Windows with 32GB RAM, and who knows what sort of NVME we're comparing on both 😂 Can you remove some RAM down to 32GB or change the NVME on the Mac to be more comparable? 😅 I'm sure your can add RAM and upgrade the Intel machine. Apples vs Oranges I only use Linux and Windows.
I have an older, previous gen M3 Pro Macbook Pro. Just ran the speedometer test WHILE watching this video and still got better score (28) than the Razer.
The .NET test actually made me believe that your razor laptop has some issues, I have intel ultra 7 1st gen and I have faster loading and runtime than your razor laptop :O
I've had a Razer Blade and loved it. This was when Apple was still using Intel chips. If I could get a MacBook that would run all of my games well, I would love to switch for the improved battery life, performance on battery, and thermal efficiency.
@revben I got an ASUS Zephyrus 4090 in 2023 so I'm not ready to upgrade yet, it's still a beast. I would definitely like to get an AMD CPU laptop in the future if I can't switch to Mac. It's unfortunate they weren't available last go around and there are only a few announced so far.
Fun video thanks. Got a Razer 16 with 4090 and rarely use it due to the fan noise. I re-pasted the GPU & CPU and VRM's to see if it'd help but no luck. The battery doesn't last either and the speakers that were meant to be ridiculously good (for Windows machines) were still pretty poor. HDR on a 4K MiniLED screen is awful too in Windows (but decent in games), although it looks pretty great on SDR. I've been banned from getting any more gaming laptops due to my fan noise complaints
You should have used a larger LLM, like llama3.2 vision 11b easily fits 4090, but being larger exploits better the huge amount of raw compute on the 4090
I wish Apple would tackle GPU rendering for 3D graphics, etc. - It's the only reason to buy Windows machines now for the most part and games, of course.
With how much windows and apple are pushing AI into their products, I would love to see this kinds of tests with dedicated linux laptops and desktops added to the roster. Seriously thinking about dropping windows and Apple products if they keep shoving AI features that you can't opt out of. Great video as always.
Mac running windows as a VM beat the Windows laptop💀💀
@ believe it or not even for engineering software it is way faster, I couldn’t believe it when I first tried and made the switch
LooooL
😶
Something wrong with that razer 😢. No way
@@rauleduardosantiestebanmor6928 have you tried matlab, cad, kicad and solidworks?
Yeah, Apple made the right decission to go their own way. Clearing bottlenecks, not just CPU speed, but the whole 9 yards
@@BeaglefreilaufKalkar efficiency
ARM architecture is doing that, not Apple per se.
@IoanBiTza Snapdragon X Elite also use the ARM architecture, but they can't even release a fanless model, because it's not enough efficient.. loosing half of the performance without a cooler..
😶
@@TamasKiss-yk4st exactly, it is apples special sauce that makes their products so superior. Ie unified memory.
Razer 18 will keep you warm during the winter season....lol
@@oscarjeong9438 apple kidoo wake up its 2025🤡
@@potataaminecraft Just wait until you got yourself a bloated battery 😅
Pros of the m4 MacBook: long battery life
Cons of the m4 MacBook: it doesn't keep me warm during the winter
@@oscarjeong9438 that’s what the PS4 is for
😂
The things I enjoy most on the MBP are the total lack of fan noise and the fact that I never have to tweak performance settings. It just runs buttery smooth plugged in or on battery. And the battery life is actually great!
I can't wait to see you review Nvidia Project Digits working with a Mac.
@@keithdow8327 Digits will run Linux, doesn't it?
@@carstenli Linux
I want to see an M4 Ultra compared against the Project Digits! I think for inference, the M4 Ultra will win. Empty your pockets Alex, it's go time ;)
@ I am a subscriber, therefore he is emptying my pockets! He is worth every dollar though.
Do it; do it! :-)
Honestly, calling the 14900HX a "brand new processor" in 2025 is borderline disingenuous. It's basically a rebranded 13th gen CPU which is more than 2 years old. Also, the RTX 4090 launched in 2022, and the RTX 50-series was recently announced.
The newly announced AMD AI Max+ 395 (yes, awful name) chip would be a much more interesting comparison. It is a "all in one" chip more similar to the Apple Silicon ones. They are expected to be decent value and power efficient too!
Yes. Amd ai 395 can have 96gb vram with 256bit memory lane.
@@passionatebeast24 You can't buy it now
can’t wait to get those in here to try
@@AZisk Yess!!! That is the matchup of the year. And NVidias Lenovo laptop when that comes out in Q4.
Hoping for the HP g1a arrives soon!
@@AZiskI really think there is some issue with the ml benchmark and no way the M4 max beats the 4090 in machine learning
Nobody is interested in Intel laptop, please test a laptop with the AMD AI MAX+
yeah, can't wait for proper iGPU tests
AMD integrated has extremely poor power management in Linux. I have 780M.
The only reason I would even consider Intel would be if I wanted TB5 to run those new eGPU's.
Other than that, Strix Halo is the clear choice.
There is no way to compare the AMD AI Max+ 395 Thunderbolt 5 speed because it won't have a Thunderbolt 5. It is impossible it has only 16 PCIe lanes.
@@Berecutecu that was the whole point I made, the only reason you might even consider Intel now is if you NEED TB5.
Before Strix Halo, it could brute force multi-thread stuff slightly better than AMD because it had more cores... that's no longer the case.
There's something screwy about that Razor. I ran Llama3.2 on my ancient RTX2080 laptop and I got 123 tokens/s. No way can a 4090 only do a 1-2% increase over a card that's two generations older and a lot less powerful. And that multicore score on Geekbench is abysmal.
What wattage does your laptop feed your laptop GPU vs the wattage fed to the Raze? Maybe that makes a noticeable difference?
😮
With my RTX4080 I got only 73 tokens/s :D
You have to set it to gaming or creative mode in the razer software 🎉
Also the write-speeds of the SSDs are not ok - I have high-end and budget m.2 SSD here and none of them breaks down on write-speed nearly as much.
The reason that both the M4 Pro and M4 Max were similar in performance may be due to the Neural Engine being identical in both chips. If properly managed, most of the network, if not all, should run on the ANE, thus providing similar results. It would be very interesting to see where the model runs during inference with Asitop (which shows CPU, GPU and ANE usage). Great comparison!
Just to be clear, they weren’t similar in performance. Using MLX-based runtimes, the M4 Max achieved 172 tokens/second to 103 tokens/second for the M4 Pro. The difference is most likely due to memory bandwidth.
Last time I checked, LLMs were not able to be run on the ANE. Maybe that changed by now, but there was something about Apple not providing 3rd party devs the necessary APIs to use it in full.
@ There is no limitation preventing LLMs from running on the ANE, it comes down to the architecture implementation and quantization. What is interesting though is that MLX, currently does not support the ANE, at least according to python MLX package on PYPI. ANE can be fully utilized though using CoreML, even for 3rd parties (e.g. coremltools on python). On my models, I can see significant speed boosts when running on ANE with CoreML.
I’ve always been a Windows user, rocking an RTX 1070m 3080m, 3090. But when the time came to choose between the Lenovo Legion Pro 7 (AMD Ryzen 9 7940HX or Intel Core i9-14900HX) for $3,400, or the MacBook Pro 16" M4 Max (48GB RAM, 1TB SSD) for $3,229 at my college bookstore, I went with the MacBook.
At first, I was skeptical I'd always thought Apple products were overpriced and overrated, especially based on things like AirPods Max and every flashy new iPhone. But oh boy, was I wrong. This MacBook is a powerhouse. It handles my 4K and even 8K drone footage in Final Cut Pro effortlessly, staying almost ice-cold and performing incredibly fast. Whether it’s 3D modeling or programming in IntelliJ, this machine is a beast.
As a gamer with over 300 games on Steam, I was initially concerned about the limited macOS library with only half of all my games available, but I’ve been pleasantly surprised. Here are the games I’ve run on 4K ultra settings so far:
- **Metro Exodus**: 100-120 FPS
- **Dying Light**: 120 FPS (200+ FPS with VSync off)
- **Hogwarts Legacy** (via Crossover): 90-100 FPS
- **Elite Dangerous** (via Parallels Desktop): 70 FPS
- **No Man's Sky**: 120 FPS
- **Baldur’s Gate 3**: 100-120 FPS
- **The Witcher 3: Wild Hunt** (via Crossover): 90-100 FPS
And there’s more!
What’s impressed me the most is how quiet the MacBook stays. Even under heavy loads, the fans barely spin up, and at times, they don’t spin at all. The laptop remains completely silent with no FPS drops. Whatever Apple is doing with their hardware and software integration, they’re absolutely on the right track-and I’m confident it’s only going to get better from here.
I have zero regrets about making the switch.
Can't wait for you to test the recently announced NVIDIA Project DIGITS!!
May!!!
@@AZisk Please test/buy two of them!!!
TWO!!?? It's already $3k for just one!!
@@AZisk could you please explain in one of your future videos about FP4 vs FP16, because if I understand correct NVIDIA Project DIGITS is around 1000 TOPS vs RTX 5090 which is 3.1 TOPS so the 5090 is around 3 Petaflobs? (BTW I'm not native English speaker I'm sorry if any mistake)
@@AZisk
As non developer, I wonder what's the point of high-end windows laptop except for gaming.
@@Dominus_Potatus just for %4 of things that cant made in mac but % is getting down.
Engineering (CAD, architecture apps, etc), data science (GIS), & many more are not compatible with macOS
You can increase ram to 192gb and 3ssd slots inbuilt. Easy to repair.
@@AzVfL Matlab works on mac. Most CAD software works in a VM.
There is no point, game set match...
Great comparison. But can you please test against a new x86 CPU like the Ryzen AI Max+ 395
When they come out, yes
@@AZisk I am waiting
Hi Alex, did you try changing the performance modes in the razer synapse software? That normally controls the tdp for both the GPU and CPU on Razer laptops. I would also recommend changing it to discrete only mode for the GPU as it would run the tests on the 4090
Did you hear the fans? I think this green beast took 3 times more energy than the apple.
Glad you're back Alex! We need you to tell us how AI is AI-ing on a daily basis
good to be back
You might want to run the same benchmark on the Razer with Linux installed. In my own experience, it's the OS that makes a lot of difference.
Have you even seen that Windows inside the virtual machine performed faster than Windows on bare metal?
Razer blade with nvidia and intel and linux is not fun. This combination will took a lot of hours for bug fixing and workarounds.
I have MBP and an old 4 year old Ryzen linux laptop, the Windows VM on Linux boots my dev env app in 5 seconds, the MBP with ARM Windows boots my env in almost 58 seconds and that's using paid Parallels...I can't stand my mbp from a dev perspective. The way keyboard "shortcuts" are implemented sucks as it's not always supported in all apps.
I've bought two 15 inch Razer laptops, 2020 and 2021 Advanced models, both had suboptimal cooling solutions for the compnents they pack and constantly overheated. The other bug bear was needing to run their Razer Synapse software in the background to be able to control keyboard lighting, fan speeds and performance modes. This coupled with their poor build & design quality I honestly can't say I would recommend them.
Poor build quality? Razer has some of the best build quality of any Windows laptop maker. I have a RB14 2021 and it's still going strong, no issues.
@@migueljardim8177 I owned a 2022 model and the edges of this laptop was way to sharp and the keyboard is one of the worst i ever used. Now its sold and replaced with a ThinkPad.
@@migueljardim8177well, I had a 2020 razer blade 15 and had the same issues that @3amael mentions
@@migueljardim8177 Don't look inside!
@@migueljardim8177 I've had to open both to resolve multiple issues and I was not impressed at all! And don't get me started on trying to find replacement components like fans, keyboards or touchpads when they die!
As much as I'm a Windows user, I love that Apple silicon is pushing Intel, AMD and NVidia to do better. Competition is good for the consumer.
Nobody needs to tell me windows is shit and I’ve even never used Mac.
A couple of things to note for ML benchmark: first I don’t think LM studio works good for nvidia chips. Usually we use vllm or sglang to serve the model which is way faster than LM studio. And the benchmarks better to be done with bigger model to exhaust power of 4090 compared to just 3B model considering it’s already 4-bit quantised. Second thing is about inference throughput. Apple mlx has never done any good batch LLM inference, meaning if you have multiple queries, it’s going to be slow to process. I can easily reach 2k tokens/second for a mobile 4090 for a 3B model with batch inference.
Yeah - I'm serving myself a local LLM on my homeserver, and while setting things up I learned that llama.cpp is really not that fast or efficient when fully loading models up into the VRAM. Things like vllm and exllamav2 with tabby or aphrodite-engine were recommended instead.
The goal is not to provide a balanced analysis of what the machines can potentially do if you use them properly. It all about shitting on PCs to make the Mac look as the superior choice no matter what. This is to lower the cognitive dissonance associated with buying very expensive hardware that isn't that much better in the real world.
Hence the sensationalist titles and whatnot. This guy keeps showing up in my recommended, yet every time his approach makes no sense appart from just shitting on PCs.
To be clear I own Mac/iPhone/aWatch. There is some good thing about it but the value proposition is out of whack...
@@clementcollier8432 but both fall under the category of very expensive hardware. It's hardly just about justifying the price.
Gonna be honest: I don't like the MacOS. I don't even have legitimate reasons, I just dislike using it.
But I do think that as things stand now, Macbooks are better for most people at least as general purpose laptops, but also for some very specific work tasks(video/foto editing, coding, audio work etc). If you need the best GPU on the market and CUDA accelerated applications, obviously that is only present on the Windows side. But looking at the device as a whole, it just looks like a horrible proposition to me - it's loud, it uses a ton of power as it's inefficient, it drops in performance and by a significant margin when used on battery (which btw was very generous of him to use it plugged in basically the whole time), consequently have poor battery life. Granted, it's a poor hardware choice and I would like to see a comparable AMD based device with some of their upcoming offerings, but these Intel machines are pretty bad.
"There is some good thing about it but the value proposition is out of whack"....hm, not sure this criticism can be directed towards the Mac/Macbook lineup. They are pretty on par with the Windows offerings in terms of value. Like sure they ain't cheap, but on the other side, the premium Windows devices aren't cheap either. As soon as you hit any sort of overall equivalency in terms of build quality, battery, screen quality, and somewhat comparable performance, you're basically looking at the same price range. This can be more directed towards the smaller devices, namely iPhones. And there's also the new, very cheap but capable Mac Mini now. It's a lot of PC for the price, to the extent that if you don't intend to do any sort of gaming, it excels at the price point compared to the competition which doesn't have anything to offer in the price range in that form factor at that performance level.
And because of my aforementioned dislike of MacOS, I've been following the market closely for the past year or so to see if something will pop up to satisfy my desire for a new non-MacOS laptop but so far, everything that showed any promise turned out to cost the same as a comparable Macbook Pro, but also turned out to have some sort of flaw that would just annoy me. In some cases they even cost more. I'm not in a rush and will wait to give AMD a chance to prove me wrong, but at this point the internals aren't even that much of an issue, but the whole package. The thing is, I may dislike their OS, but if I get a Macbook, I at least know what I am getting. It's a safe bet. Windows laptops are so tricky to get a good read on, since the market offerings are way too granular and the product lines are very messy for all of the manufacturers.
If only Macbooks could run CUDA apps... 😢
True
yeah. I know.
use mps in pytorch, or mlx, metal based support.
I think that's the real bottleneck for current MacBook compared to x86 laptops with NVIDIA (for AI usage of course)
😢
Apple Silicon may look to be one of the greatest ever technical moves. This is brutal
Now, trying generating an AI image using Flux. The MacBook will be crying from the pain, while the 4090 does laps around it.
Thank you so much for this comparison! Was thinking about getting the Blade this week over the MAC
Glad I could help!
Now would not be a good time to get a Blade laptop since very soon they'll be refreshed with the Nvidia 50 series GPU's and new AMD and Intel CPU's. I'd wait personally, especially for the AMD Strix Halo chips.
@@migueljardim8177 When new game laptop models come out I buy last years model marked 80% off. FWIW I don’t value AI frames as highly as rendered frames. The 5 series performance is numbers not experience. Why pay 10X as much so the card can hallucinate it is preforming???
I mean Razer has got it's own proprietary software called Razer synapse or something that lets you set it to quiet balance, performance mode etc. so.. maybe have a look there, that windows performance setting is just useless for gaming laptops
Small detail regarding ML test: compute increases approximately quadratically with number of tokens. So if the output on the two machines is not identical in length, there's a small bias. Also when you prompt again in the same context, more compute is needed.
In other words, when you output e.g. 900 tokens in one test and 1000 tokens in another test, there should be slightly more tokens per second in the former case.
the one thing i am interested for you to test is the new ai max+ 395 chip when it comes out, cause thats honestly the direct competitor to the m4
I really would love to see a comparison between the M4 Max and the 40x0 mobile for ML/Pytorch training. I am pretty sure you did a benchmark in one of your other videos, and I found it super useful. I know inference could be a proxy.. but imo.. training can behave differently
As always Alex, thanks for the video - great work as always!
Best thing to do with your razer laptop is returning it.
I'd be interested to see this test done with Linux instead of Windows (on the Razer).
Im noticing a big divide between next gen machines and everything else.
Fast RAM and unified memory I think is making a staggering difference. Plus as you mentioned storage speed and bad cpu's.
Mac is amazing.
I cant wait for the AMD AI Max+ 395, that might be the closest Mac competitor we'll see.
16 Zen 5 cores (32 threads) and 128 gb of RAM will be INSANE.
It will even destroy the 1 thing Intel was decent at (multi-core because they shoved so many cores in their silicon).
Unless you need TB5, the choice is clear.
Hi maybe it make sense to try Razer on linux?
I ran Arch linux with KDE on a razer blade model 2022. It runs but it is no pleasure. I dont recommend that.
@@herrspitz6964 what is the problem in your case?
@@Erwin_Anderson gonna go out on a limb and say it's arch btw
Now do these tests with Linux installed on the razer laptop. I'm curious what will happen
Excellent stuff. I was waiting for this.
👉🏻 I really would love to see a comparison between the Mac M4's and AMD AI Max Series machines.
yes, waiting for those
😮
There is no way to compare the AMD AI Max Thunderbolt 5 speed because it won't have a Thunderbolt 5. It is impossible it has only 16 PCIe lanes.
the problem with the razor 18 ... Intel.
I did not expect those results. Wow. I'm in the market for a new laptop for data analytics work. Going back to Mac wasn't even on my radar. Holy cow I have to rethink this now. GREAT VIDEO. Thanks for complicating my choices. Haha.
Windows on ARM: Am i Joke to you?!🤡
Windows on Mac: Nah i have de-bloated windows btw 🗿
Awesome vid as always. Keep it up bro. Cheers
Thanks, will do!
Great video, missed your uploads
Good to be back!
I can't wait to see the head to head with the M4 max chip vs the newly announced AMD AI Ryzen Max + 395. This should be a real battle.
There is no way to compare the AMD AI Max+ 395 Thunderbolt 5 speed because it won't have a Thunderbolt 5. It is impossible it has only 16 PCIe lanes.
For the drive, on the PC does it show up? Check the format drive application and if it doesn’t show up there try using ‘diskpart’ from the command line if you didn’t already. Most APFS formatting doesn’t show up on Windows unless you do a command line level diskpart
Great narration
Delighted to know differences between Mac and Intel performances
Linus from LTT would like to formally and informally protest these results .
I don't know if you have ever done this before, but I would love to see some comparisons running Linux. For some of these tests, I really wonder if Windows is a problem.
Thanks for this useful, realistic comparison.
Are you going to test the new Strix Halo process?
hopefully
@@AZisk hopefully at least some kind of competition. As much as I dislike MacOS this hardware is more than making up for OS. And so far its not even close.
@@heroofjustice3349The 395 will be very close and - /+ 10% in Gpu and Cpu... And better in AI.
@ yeah I know, at least thats what they say. Fingers crossed it will be also not jet engine like unfortunately most Windows laptops.
Any chance you throw a Linux distro on the Razor and re-run the tests? Obviously the disk performance will be better for things like that compilation test (NTFS is awful with lots of small files), but I’d be curious to see the difference in the AI benchmarks.
Oh, and if you want a real-world compilation test for .Net, why not the .Net framework itself?
Regarding the terrible write speeds on your thunderbolt 5 enclosure in windows you have to go to device manager -> disk drives -> right click on the drive -> properties -> go to the policies tab and enable high performance and enable write caching on the device. Now it should perform as expected.
Kinda weird very workload specific testing. Everyone watching this already knew that Apple silicon would be better suited to LLM work.
This is a channel dedicated to developers, so it's not weird at all. Half the video was about compiling code, and the other was about LLM testing. It has covered the bases there. Several channels focus on photography and videography that show the MacBooks easily beating Intel, so watch them if that is your thing. Otherwsie, what pro workload should be tested?
yeah bruv we are fedup watching geekbench and photoshop / figma benchmarks. Kudos to @AZisk for doing something that actually applies to developers.
@andyH_England Engineering workloads. Programs like ANSYS, Fusion 360, Matlab etc
@@andyH_EnglandIntel has sucked for the past five years. Try AMD
I think that in selecting 4 bit small models so they fit on the 4090 VRAM you are teeing up the MacBooks to run on the NPU. The reason I think this is that the Pro and Max chips performed the same. ~40 tops is very capable for any 4bit inference that will fit in memory.
For the 4090 to outperform the M4 you would need to do some 16bit and or learning. The other big advantage of the M4 is that it automatically sends instructions to the best core and they all call the same unified memory. The i9 is doing nothing but management of the 4090. For 4 bit inference the i9 might be about as good as the 4090.
Obsolete at this point. We have to wait for lunar lake and 5090
Are you sure the external drive didnt overheat, because the link speed clearly has been there?
Well , you weren't testing it for Gaming which is what the Razer Blade is all about. But still pretty startling. I think both the energy/heat envelope plus the faster (and unified) memory on the Mac is the big differentiator. Can't be the raw CPUs and GPUs, can it?
my bro, the large ram on the mac helps a ton
Really expected better from the Razor Blade 18. It's just not worth the price
and as a bonus, you don't need to be next to a power outlet to use the mac
Bro has same amount of ram as my sdd…
@@buddhaeyes Emotional damage 💀
next please do a competition macbook vs. desktop :D
I've heard about the price of the razer... what about the MacBook pro? In any case, yes, the M4 Max is a beast and clearly shows the benefits of designing the CPU/hardware and operating system in house.
In the US the top-specced 16" Max chip with 48gb of ram is also $4,000, the same price as he quoted the Razer.
@@MichaelGGarry That´s quite a bit cheaper than in Europe...
I would love to see the M4 Max fine-tune a model.
And we're not even talking about the build quality, speakers etc and ofc windows itself. This is just embarrassing. It's really only "good" for gaming and even at that it's bad, lots of noise and awful performance for the price. Especially on battery, might as well get a desktop otherwise.
The keyboard is the worst thing of this "Laptop"
Something is critically off here...
big time
I mean as you said, it's a GAMING laptop - it's primarily meant for gaming.
Which mac still can't beat 4090
It would be interesting to see how the Razer laptop does if you run linux on it.
razer might be loudly, not great battery, speaker and etc. but I always choices razer or other rather than apple.
I own a Blade 16 from last year and I am not offended, Alex. 😆 However...it is well known that Razer generally builds expensive machines mostly intended for gaming, not for computing LLM and string builds. As such, for a dev like you the new M4 offerings by Apple are your nirvana of course. ✌
It would be cool to see a comparison with Nvidia's AI supercomputer, Digits, which was announced at CES.
yes! as soon as i get my hands on it. looks like May
😮
14:41 yes and no, The unified memory is faster than non unified , BUT in this case , the razor 18 (I think) has sodim up gradable memory, that is slower than soldered down memory.
Doing this test right after CES is very confusing.
Something is wrong with that razer . Ain't no way it's performing like a high end snapdragon x elite 😂. I refuse to believe this
Becuase he chose tests that would favor the Max. Even a $1000 Lunar Lake Laptop would do better in these tests.
@@revben true
Windows devices are being held back by their OS and windows defender, before running any task it just starts first then lets the app to run which means extremely slow speed compared to other OS. If you run this device on Linux results will be a lot different
I would like to see some image generation tests with Stable Diffusion.
I can't wait for Strix Halo chips to hit the market, I want to see how the AI Max 395+ (naming scheme sucks) compares to the M4.
Both are SoC's that have integrated RAM and all that, so it should be a more apples to apples comparison.
I always appreciate and enjoy your reviews
New Intel Lenovos have insane fan noise too. Just connected to a monitor and running WSL my work Lenovo sounds like it is going to take flight.
I think those PC manufacturers should bundle a noise reduction ear muffs for every new Intel laptop purchased.
@@everlasts It is so bad when our sales people are calling customers the customers can hear it through the phone.
Override the fan curve. They are way to aggressive on default. My fans are off til 50C and even at 80C they are just qt 60% speed.
@@herrspitz6964 Yeah that's not going to happen on an enterprise machine.
I have had some issues with consistency from a benchmark called silverbench but it would be cool to add it to your testing because it can be run on computers on demo via any web browser. That or any other benchmark in the browser.
There's no link for your enclosure... (the one in 6:00)
Awesome comparison. Thanks!
Interesting video, but since we have thunderbolt 5 here. Can you do eGPU test with thunderbolt 5?
even if you go to settings and turn on best performance, this is still going to be based off of the power plan settings set in the ctrl panel, go to cmd prompt and turn on ultimate performance power plan, and please post the scores from there, I'd like to see those scores as well.
i can't wait for the m4 MacBook air.
The fan noise on that Razer laptop alone is a deal-breaker for me. My Mac programming needs are simple (C++, web) so I use a 15" M2 MacBook Air with 24GB that I purchased on launch day in June 2023. For me the best features are the 15" screen, the travel size and weight, the ergonomics - large trackpad and palm rests, and ... NO FANS. I think that last one puts the MacBook Air in a class by itself. Are there any Windows laptops without fans? If there were, that would be an interesting comparison.
I heard the Mac is the best to run web servers and extreme computing. Mac runs the majority of the Internet backend, right? Incredible 😊
valid point, neither do windows lol
@Harshcodes2 true, windows runs a minority is internet servers. But Windows does run the majority of corporate servers. I recognize Macs are good machines for end users in some scenarios.
Please use larger models with 14B such as qwen2.5 14B. These small models are too small to get a meaningful result also another important metric is prompt ingestion t/s. how fast it reads the prompt. A 3 words prompt is too small to get a meaningful response. Basically please do more meaningful, real world LLM tests.
Yes, both please, bigger model and longer input prompt...like 8k tokens or even more
11:35 Mac has 128 GB RAM vs Windows with 32GB RAM, and who knows what sort of NVME we're comparing on both 😂
Can you remove some RAM down to 32GB or change the NVME on the Mac to be more comparable? 😅
I'm sure your can add RAM and upgrade the Intel machine.
Apples vs Oranges
I only use Linux and Windows.
I use Linux, Windows and macOS. All of them are quite useful in different scenarios.
Would be really interested to see how the razer stands up if you do everything in a linux distro
I'd love to see the 14" vs 16" m4 max in LM Studio, I recently bought the spec'd out 14", and I wonder if I'm losing out on LM performance vs a 16"
I have an older, previous gen M3 Pro Macbook Pro. Just ran the speedometer test WHILE watching this video and still got better score (28) than the Razer.
Isn't this the 2024 Blade 18?
Hi Alex, it would be cool to see a comparison between the same laptops but with Linux in the razer one for a unix-to-unix comparison
I hope you can compare AMD Ryzen 395 vs M4 vs M4 pro and vs M4 max. that would be awesome.
There is no way to compare the AMD AI Max+ 395 Thunderbolt 5 speed because it won't have a Thunderbolt 5. It is impossible it has only 16 PCIe lanes.
I have a 14inch MacBook pro from 2021, it has the m1 pro chip, and it's the best computer I've ever had, I love it so much
The .NET test actually made me believe that your razor laptop has some issues, I have intel ultra 7 1st gen and I have faster loading and runtime than your razor laptop :O
I've had a Razer Blade and loved it. This was when Apple was still using Intel chips. If I could get a MacBook that would run all of my games well, I would love to switch for the improved battery life, performance on battery, and thermal efficiency.
So get the 395 max in April... Even Max Tech, who is the biggest Apple person, loves it.
@revben I got an ASUS Zephyrus 4090 in 2023 so I'm not ready to upgrade yet, it's still a beast. I would definitely like to get an AMD CPU laptop in the future if I can't switch to Mac. It's unfortunate they weren't available last go around and there are only a few announced so far.
Nice work. A
But aren’t you shooting on the ambulance, Alex? 🤣
Fun video thanks. Got a Razer 16 with 4090 and rarely use it due to the fan noise. I re-pasted the GPU & CPU and VRM's to see if it'd help but no luck. The battery doesn't last either and the speakers that were meant to be ridiculously good (for Windows machines) were still pretty poor. HDR on a 4K MiniLED screen is awful too in Windows (but decent in games), although it looks pretty great on SDR. I've been banned from getting any more gaming laptops due to my fan noise complaints
MS Edge running speedometer3 on Macbook Air M2 scored 20.1. And Edge on MacOS is on Rosetta. This Razer is sad🤦♂
You should have used a larger LLM, like llama3.2 vision 11b easily fits 4090, but being larger exploits better the huge amount of raw compute on the 4090
I wish Apple would tackle GPU rendering for 3D graphics, etc. - It's the only reason to buy Windows machines now for the most part and games, of course.
With how much windows and apple are pushing AI into their products, I would love to see this kinds of tests with dedicated linux laptops and desktops added to the roster. Seriously thinking about dropping windows and Apple products if they keep shoving AI features that you can't opt out of. Great video as always.