RTX 3080 recompiling issue in Matlab 2020a

Hi,

I set new GPU RTX 3080 in my PC. While trainning my deep learning nework. A warning appears and then long wait. My Matlab version is R2020a. Matlab don't support this GPU model? or is there any other issue which I should do for normal training of my network. Thank you.

CUDA version is 10.1.

The warning is:

Warning: The CUDA driver must recompile the GPU libraries because your device is more recent than the libraries.
Recompiling can take several minutes. Learn more. 

Answer 1

[Ameer Hamza]

The new GPU series by NVIDIA is powered by Ampere Architecture. MATLAB still does not support this GPU properly. You will notice unexpected behavior for now: https://www.mathworks.com/help/parallel-computing/gpu-support-by-release.html

[Walter Roberson]

Note that it is expected that there are problems with the recompiling of the Ampere series; if I understand correctly, there are bugs in the NVIDIA supplied libraries.

[Tahir Mahmood]

In general, how much time Matlab corporation take to fix compatablility issues? Thank you.

[Ameer Hamza]

As Walter mentioned that there are problems with the NVIDIA supplied libraries. You will wait for NVIDIA to fix this.

[Walter Roberson]

If the time-frames are similar to the past, R2021a would be expected to have full support for the Ampere devices.

Typically NVIDIA officially releases hardware quite close to Mathworks being about to issue a new release. For example the RTX 3080 was released to the public on September 17, 2020, whereas R2020b was released on September 16, 2020. Sometimes NVIDIA releases to the public quite late in August or early September; Mathworks most often releases on the Wednesday before the fall equinox, with the software having been out for beta testing for several months.

NVIDIA then has bugs that have to be fixed... though it sounds like they have a few more bugs than typical this time.

With the time for Mathworks to work up appropriate interfaces and do appropriate testing, and get the software to beta-testers, Mathworks would not typically have a compatible release for a new hardware line until spring; if the problems were especially bad, possibly not until the fall release after that.

Answer 2

[Joss Knight]

Does MATLAB support NVIDIA Ampere cards for GPU computation?

You should follow the advice on the GPU support by release page carefully, particularly with respect to setting your CUDA_CACHE_MAXSIZE environment variable. This ensures that you only see a single compilation delay rather than it occurring each time you run MATLAB.

Current testing shows that most functions in R2020b (or 20a) work correctly on Ampere cards although there is some incorrect behaviour for convolutional neural networks. Performance is somewhat reduced.

[Behrus Puladi]

Hello Mr. Knight,

what kind of malfunctions are to be expected regarding CNN exactly?

Thanks a lot!

[Joss Knight]

The NVIDIA cuDNN library MATLAB uses to compute convolutions returns wrong answers under certain conditions dependent on a variety of factors (such as input size, GPU device etc). Suffice it to say, it is not safe to use deep learning on an Ampere card with the current version of MATLAB R2020b.

[Jast]

Joss,

When you say not safe, what exactly do you mean by that?

Or to put it in another perspective: when do you expect Ampere to be supported? R2021a? When do you expect the release of R2021a tp happen?

[Walter Roberson]

Mathworks seldom updates functionality significantly in-between releases. Mathworks is expecting to release an R2020b update that supports Apple M1, which is a break with tradition, so I cannot say that definitely support for the Ampere line will have to wait for at least R2021a... but that would be much more typical of how Mathworks does business.

MATLAB R20xxA releases are typically the Wednesday before the vernal equinox (that is, a few days before the start of Spring), so most typical would be March 17, 2021. But that has been known to vary a little, so it might wait until Friday March 19, 2021. (Do you need the time of day? That varies more, but is typically 17:00 Eastern time for R20xxA releases, but it can take a few hours for the site to be update.)

When you say not safe, what exactly do you mean by that?

I do not know what Joss means by "safe", but the reports I see are that the NVIDIA-supplied libraries can crash and can cause wrong answers to be returned.

Reminder: Mathworks relies on NVIDIA to supply correct high-speed libraries, and does it consider it appropriate to spend their times writing their own high-speed libraries -- so if NVIDIA is supplying buggy libraries then do not expect that Mathworks will step in and write replacements that are efficient and bug-free and delivered yesterday.

[Joss Knight]

Thanks Walter.

By 'safe', I mean if you're doing Deep Learning you might get wrong answers and you won't necessarily be able to tell exactly when that's going to happen. However, if everything you're doing seems to work and you're not distributing your code or doing anything safely critical, I can't stop you going ahead and using it.

I expect Ampere to be natively supported in R2021a, and a workaround for the forward compatibility issues with Deep Learning is not out of the question for a between-release update in R2020b, if we can work out how to do it. Watch this space.