Question Passing up on AMD when they have the best product (ICC GenuineIntel shenanigans)

[SammichPG]

Hi everyone, my boss just put me in charge of buying our new workstation for image processing and deep learning.

Given articles like this one where AMD incurs a 2-3X performance penalty I am questioning the opportunity to choose an AMD configuration. https://www.pugetsystems.com/labs/h...for-Python-Numpy-And-Other-Applications-1637/

We are not bound by perf/W considerations since it would be a single machine, but we use several commercial packages (Amira, Imaris, Matlab) that we have no control over and they are all compiled with ICC or other Intel supplied math libraries.

Mine is a tiny niche (image processing of microscopy data) and we lack any ability to custom compile things to work around Intel but it makes me sick hat I will have to reward the unethical business practices of a big conglomerate...

 

[tamz_msc]

For your use case Cascade Lake X seems like an appropriate choice. Just remember to get an RTX or Quadro(if you need ECC) GPU. Even though the Cascade Lake X CPUs come with DL Boost, they're no match for the compute power of a tensor-core enabled GPU in training. Inferencing is a different story though, depending on the dataset a CPU can sometimes give perfectly adequate performance in inferencing.

 

[Topweasel]

Personal machines I would find an alternative solution. But it's not your money, you have to do right by the tool chain and if Matlab and the like are using ICC your users need to be using the best equipment for them.

What I would do is reach out to contacts in the companies and suggest that their choices to use ICC in the future might impact your use of their tools in the future. Maybe they shrug it off maybe, they don't. But its probably our jobs to put pressure on these companies to fix their tools.

 

[TheGiant]

this is a use case where the time to solve optimisation for non intel CPU takes more time than you gain from it
buy cascade lake X or xeon W
I dont know what is the licensing model (per thread or?) so make a balance of freq-thread before buying

 

[NedFlanders1976]

@SammichPG

Zebrafish scientist? We are in the exact same situation. Exact same software packages. Difference, we already tested all the software packages on AMD and did benchmarking.

With the exception of Matlab, AMD systems run perfectly fine and very fast. In particular, Imaris runs outstandingly well. Same is true vor Amira, although I haven't tested this one as intensively. Also, the Zeiss "Zen" Software works very well (not to be mixed up with AMD architecture)

So regarding Matlab. Matlab was a major disappointment at first, which is (if you did not realize so far) why I worked myself into the problem, identified the workaround and published it on Reddit. A redditor there yesterday pointed me towards your problem.

This is really super easy to apply. You dont even need admin rights although setting the variable is the solution of choice. I am also in contact with Mathworks since and they are working hard to provide an official solution in the next major releases.

I benchmarked Matlab with and without workaround on several systems (inkluding

6700k

8700

9900k

7820x

7960x

2600x

3700x

3900x

3950x

3960x

), and have the data and can provide it to you if you are interested.

Another observation with Matlab and Image processing. We are phase syncing SPIM datasets of zebrafish hearts using Matlab, which works very well. However, during benchmarking, it became obvious, that IO is a bottleneck. We decided to go for AMD with a PCIe Gen4 SSD which should be the optimal solution in terms of performance.

So ... PM me, I can send you the Matlab benchmark results. I am planning on publishing them on reddit anyway, but it will take another month until I have time to do so.

Best of luck with your science!

Ned

P.S.: Benchmark Data example [AVX2 enabled on AMD]

 

[SammichPG]

@SammichPG

Zebrafish scientist? We are in the exact same situation. Exact same software packages. Difference, we already tested all the software packages on AMD and did benchmarking.

With the exception of Matlab, AMD systems run perfectly fine and very fast. In particular, Imaris runs outstandingly well. Same is true vor Amira, although I haven't tested this one as intensively. Also, the Zeiss "Zen" Software works very well (not to be mixed up with AMD architecture)

So regarding Matlab. Matlab was a major disappointment at first, which is (if you did not realize so far) why I worked myself into the problem, identified the workaround and published it on Reddit. A redditor there yesterday pointed me towards your problem.

This is really super easy to apply. You dont even need admin rights although setting the variable is the solution of choice. I am also in contact with Mathworks since and they are working hard to provide an official solution in the next major releases.

I benchmarked Matlab with and without workaround on several systems (inkluding

6700k

8700

9900k

7820x

7960x

2600x

3700x

3900x

3950x

3960x

), and have the data and can provide it to you if you are interested.

Another observation with Matlab and Image processing. We are phase syncing SPIM datasets of zebrafish hearts using Matlab, which works very well. However, during benchmarking, it became obvious, that IO is a bottleneck. We decided to go for AMD with a PCIe Gen4 SSD which should be the optimal solution in terms of performance.

So ... PM me, I can send you the Matlab benchmark results. I am planning on publishing them on reddit anyway, but it will take another month until I have time to do so.

Best of luck with your science!

Ned

P.S.: Benchmark Data example [AVX2 enabled on AMD]

View attachment 14501

Hi Ned, thanks for the interesting post. We work on diabetic kidney disease using intravital two photon microscopy on murine models. Your experiment sounds extremely cool BTW (holy crap you are I/O limited by pcie3 ssds!), on the other hand we get to deal with the low transparency and light scattering of the kidney.
This would be an offline image analysis workstation since the instrument is located outside our office building and data would be shuttled on external hard drives.

I read about that debug flag and while a being it a fix for now, I must consider the following:
- Intel could disable the flag at any time leaving me with severely degraded performance
- there could be untested and unpredictable behaviors when using AMD's implementation of AVX2 (can you trust Intel not to screw it on purpose by exploiting some edge case of their implementation?), in science you need reliable tools and your reputation depends on them
- any time I spend dealing with issues is time I'm not doing productive things for my PhD
- any issue with the machine would reflect poorly on me since I am advising my supervisor on what to purchase (again reputation is a consideration)

I really hope that AMD keeps making good enough products for years to come in order to change the tide, contrary to other niches of IT we are very slow to change pace. It is important to raise the issue and make software vendors aware of Intel's bullshit on the cpu front. On the other hand I hope that they will make it big in deep learning with their upcoming GPUs, installing CUDA is just painful and Nvidia's prices are a racket.

 

[NedFlanders1976]

Nice project! I guess deep intravital imaging is a challenge in mouse. You should get the ins-/- mutant and simply image that in transparent zebrafish ;-)

My best budy at the Pasteur actually does your kind of stuff too. Actually with two microscopes in parallel imaging activity in both brain hemispheres at the same time. He uses a Threadripper and CUDA for analysis and processing.

Well, enough off topic. I guess people here get upset at somepoint about it.

As said, you might buy an inferior solution for more money. I did not experience a single problem or was reported a problem so far and it certainly was picked up by a large crowd (I did not expect that to happen actually). And you do not know what little present the next Intel microcode update brings to your new workstation.

That all being said, of course I understand your considerations and of course they are valid. I cannot give you any guarantee from my side that Intel does not pull the plug next MKL release or that Mathworks gets BLIS or OpenBLAS implemented in 2020a/b. And even if the solution is inferior and/or more expensive, it will still work well, so its a good solution after all.

Again, best of luck with your experiments!

Ned

 

[SammichPG]

Nice project! I guess deep intravital imaging is a challenge in mouse. You should get the ins-/- mutant and simply image that in transparent zebrafish ;-)

My best budy at the Pasteur actually does your kind of stuff too. Actually with two microscopes in parallel imaging activity in both brain hemispheres at the same time. He uses a Threadripper and CUDA for analysis and processing.

Well, enough off topic. I guess people here get upset at somepoint about it.

As said, you might buy an inferior solution for more money. I did not experience a single problem or was reported a problem so far and it certainly was picked up by a large crowd (I did not expect that to happen actually). And you do not know what little present the next Intel microcode update brings to your new workstation.

That all being said, of course I understand your considerations and of course they are valid. I cannot give you any guarantee from my side that Intel does not pull the plug next MKL release or that Mathworks gets BLIS or OpenBLAS implemented in 2020a/b. And even if the solution is inferior and/or more expensive, it will still work well, so its a good solution after all.

Again, best of luck with your experiments!

Ned

Murine kidneys have more "ISA extensions" that zebra fish "kidneys" though.

I will have to talk to my boss and set some budget expectations, she told me that there were no issues about money but we are in a Scandinavian country and the end price for a xeon W-3275 is 6500$... if she's fine with spending 50k then I could be generous with ram and buy a couple of quadros too

This biomedical off-topic was pretty wild for this forum, nice talking to you and good luck with your research as well!

 

[Gideon]

Well it looks that the fix is now finally officially included in matlab:

https://www.reddit.com/r/matlab/comments/frrhnv

 

[DrMrLordX]

Interesting. AMD is starting to get some clout.