Does the GPU version of Yambo can be run on the GPUs with NVIDIA Ada Lovelace/Hopper architecture, such as RTX 4090?

[Dean]

Dear all,
I want to perform GW-BSE calculations by Yambo on GPUs.
I wander,
1. Does the GPU version of Yambo can be run on the GPUs with NVIDIA Ada Lovelace/Hopper architecture, such as RTX 4090?
2. Which parameters of GPUs determain the performance of Yambo, such as FP64, FP32, FP16, memory?
3. Is there performance tests of Yambo on GPUs for reference? such as A40, A30, RTX6000/5000 or RTX3090/4090?
Thanks in advance.
Best,

[andrea.ferretti]

Dear Dean,
thanks for writing.

Concerning your questions:

1. I expect Yambo to be compatible with recent NVIDIA cards, including Ada Lovelace and Hopper, though we haven't tried yet (of course you need to specify the correct compute-capability on compilation, etc)
2. Yambo exploits a mixture of single and double precision computation, so FP64 and FP32 performance are the relevant ones. At present we do not use further reduced precision. Concerning memory: this is quite useful, since Yambo may be rather memory hungry
3. We have a bulk of data, the most recent performed on different deployments of nvidia A100. Results are available here:
   https://gitlab.com/max-centre/benchmarks/-/tree/master

hth
Andrea

[Dean]

Dear Andrea,
Thanks for your reply very much.
I want to install Yambo on HPC including RTX 4090 with NVIDIA Ada Lovelace architecture?
So, how to set this parameter, --enable-cuda="cuda-version,card-version"?
Any one who compile Yambo successfuly on GPUs with NVIDIA Ada Lovelace/Hopper architecture give me some instructions pleases.
Thanks in advance!

[andrea.ferretti]

Dear Dean,

in order to properly configure yambo (and actually any other code) on GPUs, you need to know the version of your CUDA library
(try eg to see in your environment variables, env | grep CUDA, os simply check which CUDA version is loaded with your nvhpc compiler)
To make an example, say you are using coda 11.7 (the one I have on my machine now).
Note that if you mismatch the cuda version used by yambo with that used by the compiler, you typically get a non-working binary.

The second number you need is the "compute capability" of the gpu: just google it with the name of the card.
In your case I get
cc = 8.9
(in my case, I have a Volta card and I get cc=70)

putting the two things together you have:
--enable-cuda=cuda11.7,cc8.9

hope it helps
Andrea

[Dean]

Dear Andrea,
Thanks for you reply.
Best，