AMD FirePro S10000: Dual-GPU Server Graphics Card

AMD FirePro S10000

AMD has launched the FirePro S10000, a new dual-GPU graphics card for professional applications and HPC (High Performance Computing).

The AMD FirePro S10000 is the first professional-grade card to exceed one teraFLOPS (TFLOPS) of double-precision floating-point performance, helping to ensure optimal efficiency for HPC calculations1. It is also the first ultra high-end card that brings an unprecedented 5.91 TFLOPS of peak single-precision and 1.48 TFLOPS of double-precision floating-point calculations2. This performance ensures the fastest possible data processing speeds for professionals working with large amounts of information. In addition to HPC, the FirePro S10000 is also ideal for virtual desktop infrastructure (VDI) and workstation graphics deployments.

The FirePro S10000 is based on two Tahiti PRO GPUs, I think the same kind of GPU that powers the Radeon HD 7950. The s10000 comes with 3584 stream processors (2 x 1792) and features a floating-point peak performance of 5.91 TFLOPS while the GPU of a Radeon HD 7950 runs at 2.86 TFLOPS. The S10000 has 6GB of GDDR5 graphics memory (3GB per GPU) and has a TDP of 375W.

The card supports OpenGL 4.2, OpenCL 1.2, VMware and Citrix direct GPU passthrough. The FirePro S10000 is priced at USD $3599.

Related links:


  • Przemysław Lib

    It will support OpenGL 4.3 when AMD will find their collective code base and add it… 😉

  • mirror

    at 375watt and offering 1.48teraflops is actually much worse than k20 which at 225watt offers 1.31 teraflops. It means that for 66% more power it only offers 13% more flops. bad deal.

  • Promilus

    “It means that for 66% more power it only offers 13% more flops”
    Not quite.
    1. We don’t know anything about GK110 efficiency WITHOUT Dynamic Parallelism – It might be just as well on par with older keplers and that’s a lot worse than GCN.
    2. Look exactly what’s that number – double precision floating point operations per second. It IS important but no less is INTEGER, BIT OPS, floating point single precision – and with all that FirePro wins.
    3. FirePro isn’t actually designed as Tesla competitor, it’s something between Quadro and Tesla. It hasn’t any (global mem) error correction, only (afaik) internal mem ecc (cache, not scratchpad/lds).

  • mirror

    1- who cares about dyamic paralellism or not. the end result matters. firepro is much more inneficient in performance/watt

    2- double precision performance is the most important factor as that’s what most of industry uses for the vast majority of the problems. But hey, I’m a fair person. Let’s also look at single precision performance/watt:

    firepro: 5910/375=15,76gflops/watt
    k20: 4106/225=18,24gfkios/watt
    k20x: 4591/235=19,5gflops/watt

    what were you saying again ?

    3- how exactly “having no ecc” helps your point?

  • Promilus

    @1 “who cares about dyamic paralellism or not. the end result matters. firepro is much more inneficient in performance/watt”
    That matters because it has significant impact on end results 😉 So tesla K20 can be very efficient if used with dynamic parallelism and gpu direct or very inefficient (in certain aspects) just as regular GK100/104/106 vs GCN.
    @2 – no it is not – first gen tesla has inferior DP performance versus RV770 but that didn’t matter much. Second gen had similiar performance to Cypress and Cayman and for almost an year now much inferior to Tahiti. And guess what – that didn’t matter too.
    @3 “3- how exactly “having no ecc” helps your point?”
    I guess u didn’t understand my point. F1 car is something different than truck right? Right. Them might have similar HP engines but different usage – and so it is with tesla and firepro. Btw cypress reintroduced in 22nm would crush perf/watt K20 with ease but of course u and your silly numbers know better.

  • mirror

    you LIED in the beginning by saying “[in] floating point single precision – and with all that FirePro wins” and when i proved you completely wrong, you start talking about “but previous versions are better ..blah blah blah”.
    Are they better just like this version is “better” ?

    lol. just lol. how much amd pays you ?

    when numbers don’t agree with us, then they are “silly numbers”.

    Please, tell your employers that they should hire smarter people than you.

  • Promilus

    “you LIED in the beginning by saying”
    No I didn’t; I based on Anandtech numbers – K20 single precision performance of 3.52TFLOPS & K20x 3.95TFLOPS so with those I get:
    SP: K20 3.52TFLOPS/225W=15.5GFLOPS/W
    I rechecked that with other sources and it seems that DP performance is really one third of SP on GK110.
    So where did I lie?
    ” you start talking about”
    Not exactly that way – VLIW4 and VLIW5 was winning in “paper performance” long before kepler. VLIW4 22nm card would still win with GK110 in peak performance but that does not translate to performance in apps. And for that last – you HAVE TO use Dynamic Parallelism, Hyper-Q and GPUDirect&RDMA – otherwise all that power stays only on paper. AMD GCN and NV Fermi is much more efficient “out of the box”
    “Please, tell your employers that they should hire smarter people than you”
    Well I guess your fanboy’s club should kick you out for trying to be smarter than u are.