M2 Max vs Intel 13th gen Python Race | XPS 15 2023

[@Cygx]

would be great if you put the numbers in a bar graph so its easier to compare.

[@rishavnathpati]

+1

[@louisant1817]

Found the UX Designer (/s)

[@louisant1817]

​@@lhl of course it's more optimal to have a graph I agree with that. It's still a UX choice, hence the *joke*

[@zm7430]

It feeds into the KZbin algorithms when a whole video is watched and replayed etc. Otherwise his hard work and research will take 1 minute or so for you to consume.

[@dabdoube92]

He won't do it. That way you'll watch the whole video.

[@georgepeppas]

Great video. Although i7 its a middle performance proc for x86 team. M2 max is the best processor (in laptop) you can get from Apple (and expensive) so it would be more fair to compare it with i9-13900hx (best of intel) and 7945hx (best of AMD) (and both are cheaper than m2 max mac *in EU). This way, you would compare all the flagship of each company.

[@akin242002]

Especially if you build your own desktop. It can be done for less than $1,500 including accessories (monitor, keyboard, mouse, headphones, mouse pad).

[@Techonsapevole]

Thanks for the Linux native benchmarks, I appreciate that!

[@Techonsapevole]

Thanks for the Linux native benchmarks, can you do the Whisper test also on AMD Ryzen 7840u or another RDNA3 card ?

[@Joch0A]

Thanks for the video! I'm just wondering one thing : does your test used the Apple Neural Engine on the M2 ? Just to understand if it was pure Apple CPU vs Nvidia GPU or ANE vs GPU ;)

[@Lazzerman42]

For me personally, It comes down to - can I use the real Visual Studio 2022 or not. And since the Apple version of Vs2022 is just a faint copy of the "real thing", I have to stay on Windows a while longer. Yes yes, I know I can use VS Code - but as a 100% MS developer, Vs2022 offers a lot more than Code. For instance there are no scaffolding in Code. I did really think MS would get Visual Studio 100% on Mac this time - but alas no! Maybe Rider would work - but I have a MSDN license through my employer - so Rider I'd have to pay for myself.

[@tunkskabulungana46]

The C++ model test would be great for me personally.because am a CSE student who whats to venture into A.I and robotis software engineering and C ++ is my best language...🎉

[@augcampospt]

Hi there from Portugal, Recommendation: can you provide a grid with the times/Results side by side in the next videos :D Obrigado(Thanks)

[@tsalVlog]

I'm willing to bet if the whisper implementation supported the NPUs on the M2, the M2 would have had much closer times, if not beaten the 4070.

[@Fozzedout]

I'm interested in the compiled whisper version. I use the python version on my M1 Air for big audio streams, and it takes an age to process 😅

[@andikunar7183]

I also would be interested in whisper.cpp tests. Pls. note that llama.cpp (and probably whisper.cpp) now also can use the Apple silicon GPUs (but need a commandline switches on start to enable this: -ngl 1)

[@autoboto]

On the mobile rtx4090=4080, whisper generates text every second with the python version. So with the microphone audio input keeps up pretty good. Whisper same code on the 2018 Mac i9 16 threads takes 30-60 seconds delay for text. Unfortunately the amd vega 20 not supported. But maybe would not help significantly.

[@PatrickDunca]

Can you do a video that compares the M1/M2 unified memory to Nvidia GPUs? Can you run more instances of an LLM, or larger parameter models with better results, on the larger “VRAM” of the Macs?

[@nexusyang4832]

That's what I would like to know also. 🙂

[@brandonw1604]

I wonder how a System76 laptop would do against the Mac. Everything for the distro built for the hardware.

[@danielkik1245]

That is the kind of benchmark test I like to see. Not some !@#$ about games or Cinebench R23. Compiler test codes shows us the real power in computers. Can you do the 7940HS?

[@nothingn992]

It would be much more interesting if the tests are done with the m2-ultra and the 13gen intel i9 (13900kf or 13900ks) on desk PC not laptop. I have made similar tests on m1 ultra and 13gen 13900kf, the results turn out that the intel machines run much faster in all cases than the apple m1 ultra machine. So the conclusion is quite clear: if the performance is what we need, we should use intel desk PC. But there still some interesting things, that the PyTorch performance (CPU) can be comparable to the intel machines (desk PC not laptop) in some cases with small samples.

[@akin242002]

I came to similar results. M2 Pro Macbook laptops are faster unplugged for non-GPU tasks. Plugged in or a PC desktop, high end CPU PCs are faster. With Radeon or Nvidea GPUs, they are faster too.

[@AndriiVozniak]

I have an XPS 15 with i9-12900HK, RTX3050Ti and 32GB of dual-channel DDR5 RAM 🤝

[@AZisk]

that’s a good one. tested that last year

[@AndriiVozniak]

​@@AZiskhope you liked it. I have Win 11 & Mint 21.1 in dual boot, never tried WSL tho

[@PixelFerret]

@@AndriiVozniak WSl has great integration between linux and windows, but is limited otherwise. I fear a lot of workflows, specially those center around containers, are better on linux native. However, the nvidia drivers are problematic to say the least ( currently ). so is a mix bag. as an example, i am trying to find the actual code that @AZisk ran here since i suspect that there might be a weird issue in the results, i want to try to run them natively on windows and see what i get. This of course is for curiosity purposes as Alex has always been very thorough on his tests and videos. We are getting to a point in tech that we are saturated for options and so many different possible configurations can have an effect.

[@hderazi]

Hi Alex, what do you think of Asus ZenBook 14X with 13900h without gpu. Is it worth a buy at this time ? I use Mac Studio, but I am traveling soon.

[@mrflipflop7116]

If you deactivate real time protection on windows 11, you gain a lot of perf (and I mean a lot, specialy if you do some ML/DL training or inference).

[@helbertgascon]

I'm curious. Is the Mac running in auto mode or high power mode? I'm not sure if it's 14 or 16 inch but the 16inch m2 max has a high performance mode. Also I'm not a mac user so I don't know but the m2 is a SoC with a 26 core neural engine. Can't you utilize it to do AI/ML tasks instead of GPU?

[@oida10000]

Did you ran the tests on Windows too? I mean directly not on WSL?

[@KokkiePiet]

So what you are basically testing here is pure core speed, the programmes are only active in the CPU/SoC. No I/O to the SSD. IMHO this is a very limited test, it doesn't say a lot about real world performance when you are for instance compiling a big stack, editing video or photos, doing data mining, etc etc

[@nexusyang4832]

Given the amount of RAM your studio has, is it capable to load the larger language models since even the 3090/4090 only has at most 24GB of ram?

[@gustavodutra7548]

where can i found this benchmark's codes to compare with my machine?

[@Roman-rh2dk]

Why do you bring old asus into this test?

[@m.goedeker7381]

What about the tests on the M3 max?

[@PeterEide]

Im buying my first mac, should i buy the MBA 15?

[@blackdedo93]

is the M2 running metal?

[@mvbproductions_in]

Brother can you do video edit review! Thanks.

[@elgs1980]

We want to see M2 Ultra.

[@m.goedeker7381]

Is there also a way to leverage the apple ane for training tasks for ml?

[@papasmokyy]

could you please please show the results side by side in a table or so.... it is hard to follow sometimes

[@arozendojr]

Too difficult to use asdf on Apple Chip? Does it work installing sdk, download as x86 or ARM?

[@DanielHomeImprovement]

thx for your interesting videos. LEt me add something here comparing speeds of Mac M1/M2 vs intell/NVIDIA or AMD architectures. If you use the laptop for any kind of AI/ML at least for now you better stay with Intel and NVIDIA. I personally use Pytorch for ML and although some of its functionalities have been ported to Mac Metal most of them dont, so in a project you cant use the GPU hardware acceleration on Mac compared to NVIDIA, and we are talking about speed differences from 10 to 50 times. I know plp will say that you dont use your machine for training models or any complex models but anyway we are losing A LOT of performance on this area. HOpefully in the a few years this will change, but for the moment Mac doesnt have the level of performance that a n Intel/AMD/NVIDIA laptop provides for AI and ML

[@tomtang8708]

Exactly. It is just not worth it to risk the compatibility/optimization issue to use MBP for some specialised professional tasks, even though MBP is so comfortable to use (basically zero fan noise).

[@worksonweekends]

See if you can adjust whisper to use --device mps instead of cuda

[@atgred]

Were the PC laptops connected to power inlet? I’ve seen tests where if not connected to power, the cpu works a lot a slower.

[@AZisk]

very true. i tested with and without in the video

[@jamilal-idrus1905]

In my test, i cant use a GPU from my computer to accelerate train process in Machine Learning like classification or clustering, but while i am using deep learning like image classification or image segmentation, the GPU from my computer can accelerate train process, so what if you make test on this device apple M2 chip and Intel Gen 13 to train 5000 image, wich one will finish first in train and testing model process?

[@charlessnachez4248]

Bro just put the results above each laptops

[@shayanabbasi1668]

Hey I need your help sir so fast I'm learning python right now and I want to work so hard on web 3.0 and Ai, for These I am looking for a laptop and I don't know what should I buy, consider me as a rich person because it would be my future, so it's not important how much it Costs just tell me what you think the best Option is. Thank you sir 🤝

[@gsanto]

Hi Alex, I was wonderig if you were running whisper on Apple Metal (GPU), though. Cause you can use CUDA on Apple Metal... I was able to run several models on Apple metal but whiper. And, of course, if you were running it on CPU well, that's not a fair comparison :)

[@AZisk]

I believe this was a cpu only test (should have mentioned that). However, I’m curious about how you can use CUDA on metal. Can you elaborate? Feel free to email me

[@broccoloodle]

"Linux is slower?". Alex: "No no no 😂"

[@PixelFerret]

Alex, do you have the right drivers for the xps? i fear that you are a) using an old version of linux or b) using the integrated graphics on the windows laptops due to not having the right drivers. Nvidia drivers have been particularly bad on linux lately

[@AZisk]

yes, i have the right drivers. had to install them separately

[@PixelFerret]

@@AZisk it seems oddly slow, the RTX should have performed better, unless it is because it's the laptop model.. :/ where can i find the test that you ran? i would like to try them on my system

[@bdarla]

It seems that the M2 was running in CPU only, right? If one manages to employ the GPU for ML, then it gets a lot better. I did some training with Pytorch and using GPU is 10-11 times faster than CPU. Of course, it might not be trivial to run Whisper on M2 GPU...

[@AZisk]

yes i believe you are right. I don’t know if there is a mps version of whisper out yet

[@antonrrtyq]

1:11 shwartz who ?:)

[@anmiasm]

It would be interesting to compare power draw from different chips

[@axaysushir]

Should I buy HP Elitebook 845 G8 Ryzen 3 16GB RAM 256 SSD for programming? I want to run Docker, Android studio & other day to day tasks. Or go with a Mac Mini?

[@AZisk]

Go with Intel especially for Android dev

[@axaysushir]

@@AZisk I've ThinkPad P14s Gen 1 i7 Pro 32 GB RAM but it heats up too quickly and it slow downs laptop sometimes & Battery backup is not good. I think I should build a custom PC.

[@zhanwenchen9238]

PyTorch on M1/M2 can be a little janky.

[@gordonfreimann]

i have a feeling that it didn’t use gpu on the mac version. maybe i am wrong

[@AZisk]

i think whisper didn’t use mps, so cpu only.

[@DamianKleiman]

I've been using Windows + WSL and very happy with speed on the 13th gen intel. But it seems native Linux is quite better. Thanks for the comparison.

[@fitsuit1555]

You need to display all results side by side

[@tomtang8708]

Intel+Nvidia laptops are for full on performance, professional tasks, compatibility and gaming, even video editing 13th+rtx 40 has significant advantages in most tasks now. The main issue is the battery life and higher chance of fan noise even in light tasks, mainly due to intel's falling behind of TSMC in advanced lithography. It's exciting to see intel pushing to catch up TSMC in lithography. MBP is for efficiency, zero noise on medium load tasks and fantastic battery life, but still very poor support in those very specialised tasks like specialised engineering software and lots of research environment (like in my PhD project all the researcher either stick with intel+windows/linux or buy intel macbooks, that physics code just cannot compile on apple silicon no matter how hard you try.) Also even though apple silicon definitely can run some AAA games, the issue is that the performance is still poor and not stable (compared to the theoretical performance of apple silicon GPU), the 1%low fps is poor and there are many stutters due to non-native graphics support, so the AAA gaming experience is terrible for MBP pricing.

[@tamalchakraborty5346]

Just curious, what kind of "physics code" are we talking about here. Is it up on github?

[@yohcg]

I was wondering what is the brand of your chair? Seems like Herman miller. But not sure. Nvidia rules. But actually even if mac is slower, it is impressive.

[@AZisk]

it’s a really old herman miller aeron

[@theyaghu]

Where is that linux video that you said 4:40

[@AZisk]

coming shortly

[@theyaghu]

@@AZisk okay.....

[@Heidi0039]

I would like to see a neural engine interference using CoreML.

[@woolfel]

even though I use python, the performance is just so horrible for anything intensive. as a wrapper for C/C++ code it's fine, but too many fortune 500 companies have drunk the koolaid. I see lots of companies using python for ETL IO intensive batch jobs. The performance is just horrible, but lots of companies are doing it and paying Azure lots of $$. I'm hoping Apple keeps improving metal support for tensorflow and pytorch. Once they get the neural engine support beefed up, it should close the performance gap.

[@sonykdl8053]

Sir, I have a problem with my VS code. Every time I write #include, the VS code tells me that I should update my include path, What should I do? Please reply to me,I am from Bangladesh.I am a macbook pro m1 max user.

[@GURSIMARSINGHMIGLANI-v7j]

4 seconds difference? who cares

[@Rydn]

I though the gpu memory on the mac would be an advantage for those ML tasks.

[@drill_fiend1097]

Nvidia pretty much has more time to polish CUDA. Also, shared memory will have much more latency despite flexible size.

[@marko95g]

so xps with i9 will blow m2 max away :D xps 15 ❤ macbook 👎

[@krishnachaitanya4563]

Mac always wins ❤

[@RomPereira]

My desktop, i5-13400 with DDR4 memory is faster than my Macbook 14 M1 Pro (base model) by a mile... Windows 11 is becoming one of the best OS to develop on. Apple started a battle, but know it is falling behind. Do not get me wrong, I do love my macbook, it is an incredible notebook for the price I got it for (1599 + taxes), but a debloated install of windows 11 pro is awesome and it is kicking mac os but. Intel CPUs are really impressing me at them moment. But as a mobile unit goes, the macbook is still the champion.

[@Ti-wy1fg]

Same, I've been using intel 12th PC for development, and a M1 macbook air for on the go stuff. Apple silicon is great for laptop: light, no fan, crazy performance per watt. But for pure performance, intel 12th and 13th is top.

[@RomPereira]

@@Ti-wy1fg agree 500%. Nothing beat MacBooks on the go.

[@Rednunzio]

Intel is faster on every day task for developer ?

[@bichengchen7593]

Love your videos! A couple of comments: is numpy using Apple Silicon Veclib header. Pytorch nightly build has support for Apple Metal and mps. It will be interesting to see if that helps to speed things up.

[@andikunar7183]

Also not sure if Alex did invoke whisper on the M2 max to also use the GPU (--device=mps) - maybe check with asitop, if it fully uses the gpu.

[@jaswanthb8833]

check out the tgp as xps 15 model u r using iscapped to 40W try out something with 80,95,115W to find the difference

[@nickplennox69]

What about benchmarking the Studio M2 Max and Ultra for developers? No-one else is covering programming benchmarks (Xcode, Android Studio) on these machines! Just a suggestion. Thanks for the continued dev content.

[@AZisk]

I covered M1 Ultra last year, and the demand for it was really really low. So i’m skipping the M2 ultra this year

[@nickplennox69]

@@AZisk Thanks Alex. I imagine it's a lot of work!

[@MeinDeutschkurs]

Schwarzenegger V3.0 needs definitely thinner pipes. 😂😂😁😁❤️❤️

[@AZisk]

it should be liquid metal

[@rukawata]

i wanna see you test it in the coffee shop lol

[@worksonweekends]

Arnold can satisfy 3 women at once

[@allaboutferrets]

2:15 that’s the interesting way to create the counterweight😂

[@AZisk]

thanks. its a light mic :)

[@allaboutferrets]

@@AZisk I’ve followed your instructions on how to set up Base M2 Mac Mini with an external ssd. Did you find the work around the slowing macos during the workload?

[@BahadrCAKARhosgeldin]

What a finger system

[@alma-p]

Alex, would you recommend a 16inch m1max 64gb 1TB like new for 2800usd or an unopened 16inch m2max 32gb 1TB that cost 3250usd? I'm not sure if the CPU really makes a difference in the real world, especially considering that i could get an m1max which has 64gb and has 6cycles. Basically new and cost less. Does the m2max generally feel any faster or is it a waste of money and better just save up the extra for next years m3x?

[@endaksi_channel]

thanks for proving that "Performance" mode on Ubuntu really works!

[@kushagrasingh7731]

Great video, so what would be the best laptop for learning and practicing ML. Thanks

[@bakenbard]

can we have a little graphics at the end of each comparison? maybe concluding table or something..

[@arabporr]

Awesome 😍

[@PatrickDunca]

Considering all the noises AMD have been making about PyTorch, it would be interesting if you were able to compare one of their GPUs.

[@АлексейШилин-д1ф]

Which python versions were you using, especially on Linux machines?

[@colddust21]

didn't even get a notification for this one

[@develocode7773]

what do you do with all these laptops? besides testing

[@AZisk]

oh i think you know

[@ameliabuns4058]

I'd love a c++ vs python performance test for the AI stuff. I always wonder how much prue c++ and c helpers with python differ

[@niravlekinwala]

Can you please share the sorting code that you used in the first test? I just want to test the results on my Laptop. I think the performance of Numpy can be improved by using Apple's Accelerate Framework.

[@kristhianznr]

The XPS is like my wet dream... and the result of the OpenIA test is incredible, especially if we take into account that in the XPS, the GPU is limited to about 40 watts, if that test is run on a laptop where the 4070 Do not have voltage limitations, the result will be even more incredible. Big Win for RTX 40 Series. But the MBP for ML is another level, incredible results!

[@kristhianznr]

Also, u can do videos like XPS 15 vs MacBook Air 15" ? Although there is a difference of $550 in favor of the MBA, I think a comparison of those two would be fair.

[@AZisk]

Don’t get the XPS wet, it’s not waterproof 😂

[@AZisk]

mba vs xps vid coming

[@kristhianznr]

@@AZisk 😂😂😂

[@kristhianznr]

@@AZisk awesome! Thanks 🫡.

[@ahmetsarkis6313]

I think it will make a difference in m2 ssd literacy rates. The result can be surprising if all machines use the same m2ssd.