NVIDIA A2 performance in Matlab R2023b is lousy

We've got a new GPU: NVIDIA A2 (~1400€).

Its performance with Matlab R2023b is really bad.

Let's compare it with an old NVIDIA Geforce GTX 1050 Ti (~200€) by using gpuBench (double precision results, higher is better):

I wonder why it performs so bad. In all technical aspcets, the A2 is superior. Any idea?

Answer 1

[Joss Knight]

The A2's spec says its double precision performance is 140 GFLOPS vs the V100's 7 TFLOPS, so this is pretty much expected. Even its single precision performance is significantly worse. It also has a weak memory bandwidth.

I guess the lesson is that NVIDIA wants to give you a range of performance vs price trade-offs even with the latest architectures. I'm guessing the A2 is coming in at about 1/10th the price of the V100, or thereabouts?

https://en.wikipedia.org/wiki/List_of_Nvidia_graphics_processing_units#Tesla

I'll admit I'm less certain why it performs worse than the 1050. Obviously there are a lot of different factors involved, but you might want to contact NVIDIA. As a data center GPU it does have certain power and cooling requirements that could affect its performance if not done right.

[Matthias Wurm]

Thank you very much, Joss, for your support,

in the meantime I found out that it is definitely not a Matlab problem.

I also checked the FP64 (double-precision) performance with another tool (CUDA-Z). The results are similar to those obtained with gpuBench.

The NVidia A2 is a very poor choice for FP64 calculations, even if I could reach the nominal 140 GFlops/s, which I could not. (I will discuss this with NVidia support).

An interesting discussion on this topic can be found here.

[Joss Knight]

Very useful to know, thanks for responding.

[Matthias Wurm]

It's also not a computer/mainboard/bios-problem: In the same computer our Nvidia V100 works perfect.