RTX 3090 vs A100 in deep learning.

I ran ResNet on RTX 3090 and A100

Performance is better in RTX 3090 about 1.2 times than A100

I searched and found out that GPU Coder helps use TensorCore

So, I want to be sure if I use GPU Coder, A100's performance is going to better before I purchasing GPU Coder

Thanks.

Answer 1

[Joss Knight]

According to the spec as documented on Wikipedia, the RTX 3090 has about 2x the maximum speed at single precision than the A100, so I would expect it to be faster. The A100 is much faster in double precision than the GeForce card.

Both will be using Tensor Cores for deep learning in MATLAB.

[Kyle Lee]

I have question about Tensor Core.

How can I proceed with training using Tensor Core in deep learning?

Shoud I use GPU Coder?

If not, can you tell me how to use Tensor Core without GPU Coder?

[Joss Knight]

Tensor Cores just get used when appropriate for convolutions when you use a convolution layer or the dlconv operation. They also get use for matrix multiplication and various operations that use it, in double precision. You don't need to do anything.

[Kyle Lee]

For understanding, I ask you to confirm.

Then, If other models like Alexnet, Googlenet etc.. are used , do these models automatically use Tensor Core?

[Joss Knight]

Yes.

Answer 2

[David Willingham]

Hi,

What version of MATLAB did you test ResNet out on? I'd recommend running benchmarks on the latest version of MATLAB.

Was it for inference or training?

AS an FYI, you can contact MathWorks and receive a trial of GPU Coder to test out the performance first hand.