We have some A100s in a lab or two that have run benchmarks, but we do not
yet have official numbers. We do have Amber running on A100 with proper
compiler settings, so 3080-Ti and 3090 are not far behind (if we don't have
it working already). With no further code changes, the A100 cards were
about 30% faster than the reigning V100 for JAC (24,000 atoms), and up to
50% faster for larger systems. Very impressive results, especially
considering that the A100 requires that the code uitilize explicit
__syncwarp() commands that the earlier cards can do without (these
synchronizations amount to 5-15% of the effort if earlier cards are
required to use them). This is happening in the context of the cards
having about 30% greater fp32 instruction throughput (19.5 TFLOPS) than a
V100, so I wouldn't count on new kernel tuning to carry us much further--I
suspect the cellulose case, in fact, was benefitting from the better memory
bus as much as the higher FLOPS.
NVIDIA is once again pushing the envelope, but the real innovations with
the A100 are in the fp32 tensor cores which we do not (yet) marshall for
our purposes. The fp32 tensor operations are much more suitable for our
needs, however, so it's becoming easier to see MD consuming the 150 TFLOPS
that the A100's tensor cores provide.
One thing I would caution about, especially for people eager to buy Ampere
and feed them with existing PCI/Express cables, is that power draw is
likely to be a limiting factor. I doubt anyone is thinking about an A100
except in the context of an HGX server, which will have the electricity
problem solved out-of-the-box. Gamers are happy with the RTX-3080, saying
that it delivers 25% more performance on 10% more juice than the last big
thing, the RTX-2080Ti (this is in line with what the A100 should deliver
over the RTX-6000 or V100). Specifically, they're talking about a 300-350W
power envelope, compared to a 275-300W envelope. The A100 itself is rated
up to 400W. Take a look at the following specs:
https://www.nvidia.com/en-us/data-center/a100/
<
https://www.nvidia.com/en-us/data-center/a100/>The only difference in the
HGX and PCI/E cards? 400W in the former and 250W in the latter, which has
to mean that the latter is not going to operate near peak FLOPs for any
sustained period of time. my feeling is that, if you hook any Ampere card
up to a power supply that wasn't designed to support 350W, you're not going
to see that much more performance than you were getting with your
RTX-2080Ti or Titan-V. There are some impressive features in the Ampere
cards, but the improvement in FLOPs / Watt is incremental and, once again,
they are loaded with cores that we and other scientific programmers are not
yet taking advantage of. The cause for optimism is that it is becoming
more likely that we will take advantage of these innovations with only
moderate changes to our algorithms.
Dave
On Fri, Oct 9, 2020 at 11:37 AM Mike Mazanetz <mikem.novadatasolutions.co.uk>
wrote:
> Dear All,
>
>
>
> I would like to know whether there has been any benchmarking against the
> NVIDIA GeForce RTX 3080 GPU cards and whether there is any doubt whether
> this would work with
>
> the latest version of AMBER and previous version, eg AMBER16.
>
>
>
> Many thanks,
>
> mike
>
>
>
>
>
> _______________________________________________
> AMBER mailing list
> AMBER.ambermd.org
> http://lists.ambermd.org/mailman/listinfo/amber
>
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Fri Oct 09 2020 - 10:00:02 PDT
