On Tue, Mar 28, 2017 at 10:27 PM, Chris Neale <candrewn.gmail.com> wrote:
> Dear Jason:
>
> thank you for this awesome information! I did take a look at that benchmark
> page first, but it doesn't contain the word "nvlink" at all.... is there a
>
The very first plot has a bar titled "2X Quadro GP100 PCIE+NVLINK" which
runs at 832.28 ns/day with a 4 fs timestep and hydrogen mass partitioning.
The bar right below that is titled "2X Quadro GP100 PCIE" which runs at
747.09 ns/day with the same settings. Note that "nvlink" is in the image,
so it won't be found by a simple text search.
That gives you a ballpark ~12-14% speedup that you get from NVLink.
> synonym for nvlink or something? Sorry if this is obvious, but I've had
> very little experience with nvlink, only on a dgx which has faster GPUs
> than any other nodes I have access to so I am not sure if its the GPUs or
> the nvlink or something else (besides shared PCIe bus, which you explained
> very nicely) in the dgx that is helping the scaling.
>
NVLink is something different (more?), and came after I stopped paying as
close attention to the new developments in GPU land as I once did. It's
not P2P (since it can also operate between GPUs and CPUs). The bandwidth
is very high (80 Gbps IIRC), but I didn't see any reports about its latency
(although to get a 12% speedup over P2P suggests it can't be that bad).
At the end of the day, even with NVLINK and only 2 GPUs, the parallel
efficiency is still a tad underwhelming (although it will be better for GB
which utilizes a crappier-scaling algorithm that nonetheless is more easily
parallelized). 2 GPUs runs at only 134% the speed of 1 GPU (for the very
high-end GPUs, where the scaling is expected to be the worst).
All the best,
Jason
--
Jason M. Swails
_______________________________________________
AMBER mailing list
AMBER.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber
Received on Tue Mar 28 2017 - 21:30:03 PDT
