3090 vs A6000 for CAE FP64

[BroderSalsa]

I’m working for a small engineering firm, that produces water simulation software.

I’ve been tasked with picking parts for a new workhorse pc. My colleagues doesn’t know anything about GPU’s and I’ve only built them as a hobby for gaming, so I was the best candidate (:.

For our price range, I’ve narrowed it down to either 3090 or A6000.
As I see it, the VRAM to cost ratio is basically identical for the 3090 and A6000, and the FP64 performance to cost ratio is also somewhat identical for the two cards. If the statement above is true I’d recommend the A6000, as it has much lower TDP than 2x 3090, making it easier to cool, and add cards later on.

However, I have been unable to find any FP64 benchmarks that have tested both the A6000 and the 3090 card, to confirm the numbers in the specifications.

I would like to find a FP64 benchmark that doesn’t involve any ML/Deep Learning/AI, as I don’t want the additional tensor cores in the A6000 to impact the results, as these will be largely useless for our use case.

The use case is computing linear algebra, specifically complex matrix operations.

Right now I’m going to suggest we get 1 of each card, and test with our own software, however it would be nice to have some actual FP64 benchmarks to show.

I’ve been scouring the internet, but to no avail, if any of you has some relevant articles lying around, I’d love to read them!

[drbsg]

If FP64 performance is what you need then neither card is going to give you much over a modern high-end CPU. (Off the top of my head an AMD 3990x can do about 4 TFlops on the right application, the A6000 is listed at 1.25 TFlops peak).

There are two options you haven’t mentioned, one of which may be in budget. The expensive option is the Nvidia A100 which has full FP64 support and claims 9.7 TFlops FP64 and has 40 or 80GB RAM. The less expensive option is the Radeon Pro VII with 16GB RAM and 6.5 TFlops peak, and engineering applications is its target market.

[BroderSalsa]

Unfortunately we also need a buttload of VRAM, so the GPU suggested will not suffice. I appreciate the help.

We are already at least 20x times faster than our competitors with a pc with 4 2080 TI. The biggest problem we have is lack of VRAM. Many simulations has to be split into 30 parts, which we have to stich together in the end… So an increase in VRAM is necessary.

[nnunn]

Hi BroderSalsa, you mentioned:

@BroderSalsa

“FP64 performance to cost ratio”

If VRAM and fp64 performance are both important, then a single A100 might be worth a 2nd look. Latest PCIe variant has 80 GB HBM2e with twice the bandwidth of a 3090… and 20 times better fp64 performance.