[Beowulf] AMD cards with integrated SSD slots

[C Bergström]

On Thu, Jul 28, 2016 at 10:33 PM, Prentice Bisbal <pbisbal at pppl.gov> wrote:
>
> On 07/27/2016 05:53 PM, Brian Oborn wrote:
>
> I noticed that AMD is coming out with a new line of Pro cards that have two
> PCIe 3.0 M.2 slots for caching 1TB of content on the card.
>
> http://www.anandtech.com/show/10518/amd-announces-radeon-pro-ssg-fiji-with-m2-ssds-onboard
>
> I'm wondering what the list's thoughts are on the following questions:
>
> 1) Is AMD relevant in HPC anymore? In the Top500 list I only found two older
> systems that had AMD video cards and none that had AMD CPUs. Does anyone
> here run smaller newer AMD clusters?
>
> Are you talking CPUs or GPUs? As far as their GPUs, when AMD released their
> FirePro S9150 GPU, it was signficantly faster than NVidia's K40. I don't
> think NVidia has released anything new that can beat it yet. The K80 is
> really just 2 K40s in one package.  I haven't been keeping up with industry
> news as much recently, so NVidia could have come out with something better
> since then. AMD has since come out with the FirePro S9300 x2, which based on
> the name, I'm guessing is two S9150s in a single package, a la the K80
>
> http://www.pcworld.com/article/2462240/amd-claims-supercomputing-gpu-performance-crown-with-firepro-s9150.html
> http://www.amd.com/en-us/products/graphics/server/s9300-x2
>
> At SC15, AMD announce that they were adopting CUDA (sort of)
>
> https://www.techpowerup.com/217611/amd-launches-the-boltzmann-initiative-brings-nvidia-cuda-to-firepro
> http://wccftech.com/amd-cuda-compilercompatibility-layer-announced-with-the-boltzmann-initiative/
>
> While NVidia is still dominating the GPU field by a large margin, these
> developments show that AMD hasn't given up on the GPU market and has made
> some pretty significant advances in the past couple of years. Only time will
> tell if this is enough to keep them in the game and grab some market share,
> though
>
>
> 2) Are there many workloads that would benefit from this type of cache? I
> think trying to juggle GPU RAM, system RAM, GPU storage, system storage, and
> inter-node networking might make this too difficult to scale beyond a single
> system?

I'll let the reader decide actual performance between s9150 and K40,
but you can do pretty close side by side SPEC ACCEL comparisons
looking here
http://spec.org/accel/results/accel_acc.html
// Shameless self promotion - yes s9150 + PathScale is the #1 result

the s10000 was the 2x s9150 as you describe it (afaik)
http://www.amd.com/en-us/products/graphics/workstation/firepro-remote-graphics/s10000

Any of the newer cards are likely a bit faster or with more ram than the s9150.

The s9170 had more GPU memory and so check clocks and published Flops
for any parts which come after that.

If you're on a budget and want to test things - the w8100 (active
cooling) had performance that seemed to sit between a K20x and K40.
I've seen them for *really* good deals on newegg..