SC12 is officially here and the chip vendors are eagerly touting their latest and greatest offerings. On the GPU accelerator front, Advanced Micro Devices (AMD) and NVIDIA both have products dropping today. AMD is launching the FirePro S10000, its most powerful GPU card yet. In fact, AMD is claiming this is the most powerful server graphics card on the market, but we'll come back to that in a moment.

The FirePro S10000 (yes, four zeros!) is the successor to the S9000, which debuted in August. Both chips are based on 28 nm "Tahiti" silicon and support error-correcting code (ECC) memory, but compared to its predecessor, the S10000 has an extra GPU and packs more FLOPS punch (both single and double precision). The new card also offers greater memory bandwidth (480 GB/s versus 264 GB/s with ECC turned off) and slightly reduced clock speed (825 Mhz versus 900 Mhz). Both cards have 6GB of GDDR5 on-board memory and support PCI Express Gen3, but while the S9000 has 1 DisplayPort output, the S10000 offers 4 and a DVI. Total core count doubled: from 1,792 to 3,584. These enhancements raise the maximum heat envelope from 225 watts to a daunting 375 watts.

As with the previously-launched "Southern Islands" based-cards, the S10000 is built on top of AMD's "Graphics Core Next" (GCN) architecture, which enables the two GPUs to carry out compute and graphics processing simultaneously. This makes them a good fit for a range of visualization and technical workloads, but the target audience is design professionals who use computer aided design (CAD), and media and entertainment (M&E) applications.

At last Tuesday's press briefing, Bahman Dara, senior manager of worldwide marketing for AMD, explained that as products and design get more complex and sophisticated, there's a greater demand for computational analysis, which makes server-based computing and remote graphics increasingly important. "This product is capable of delivering both compute and graphics at the same time – why buy two cards when you can buy one and achieve same objective?" queried Dara.

The "two cards" Dara is referring to are the two NVIDIA lines: Tesla, optimized for compute, and Quadro, which is geared for high-end graphics work. The idea is that the S10000 can do the work of both of these chips. During the same briefing, Joyce Burke, product manager at AMD, continued the theme. The card's flexibility will enhance IT integration, she said, and because users will not need to purchase a second card for specific tasks, it's also cost-efficient. Speaking of cost, the S10000 retails for $3,599 US, $1,100 more than its predecessor, but, as usual, most sales will take place through OEMs.

The continuing push by AMD into higher-end GPU territory shows they are serious about competing in this market, and that means competing against NVIDIA. The chipmaker emphasized repeatedly during the press briefing that its latest graphics card outperforms NVIDIA's best offerings on pure performance and performance per watt. However, and this is important, they were using stats from the older M2090 and Kepler K10 products. We've known the higher-end Kepler K20s were coming since NVIDIA broke the news at their GPU Technology Conference last May and we knew they would exceed a teraflop of peak double-precision performance. With the full K20 specs now available, these older comparisons are obsolete.

"This will be the first professional-grade card to exceed one teraflop of double-precision performance," Dara told reporters last Tuesday. Alas, the NVIDIA K20 and the uber-premium K20X, which also dropped today, both exceed the teraflop mark as well. The Kepler K20X is capable of 3.95 teraflops single precision and 1.31 double precision, while the K20 offers 3.52 teraflops and 1.17 teraflops, respectively.

With 5.91 teraflops of peak single precision and 1.48 teraflops of peak double precision floating point performance, the dual-GPU FirePro S10000 maintains some bragging rights. But it does so at the expense of efficiency. When it comes to double-precision performance, the new FirePro runs at 3.95 gigaflops per watt, while the K20 outputs 5.2 gigaflops per watt and the K20X achieves 5.57 gigaflops per watt. On the single precision side, the figures are: FirePro (15.76 gigaflops per watt), K20 (15.64 gigaflops per watt), and K20X (16.81 gigaflops per watt). A more apt comparison, however, is to the single-precision-optimized K10, which supplies 20.35 gigaflops per watt (making it 23% more efficient than the S10000).

Burke acknowledged that the S10000's max thermal design power of 375 watts is at the high end, but emphasized that with two GPUs in a single dual-slot configuration, the new FirePro uses 15 percent less power consumption overall than two 225 watt cards. Going by AMD's pre-release press material, Burke is most likely referencing the 225-watt Tesla M2090s, but note that this is a common power envelope for a GPU accelerator. It shows up in the FirePro S9000 and in the Kepler K10s and K20s; the elite K20X will boost the heat output to 235 watts. At any rate, FLOPS per watt is a more useful metric than overall power consumption, and as we've demonstrated, these figures are not in AMD's favor.

Where the new FirePro excels is in pure FLOPS and its flexible compute-plus-graphics design. The S10000 is the Swiss Army knife of processor cards – it supports HPC workloads as an accelerator, but is also positioned and targeted for Virtual Desktop Infrastructure (VDI).

"FirePro products support two types of VDI," Burke explained. For higher-end applications, direct GPU pass-through mode from VMware and Citrix facilitates one user per GPU – and since these boards have two GPUs that means two users. Reflecting a more traditional VDI experience, Microsoft's RemoteFX allows one GPU to be shared among multiple users of office productivity apps that are not particularly graphics-intensive.

AMD's claim is that with two GPUs per board, the S10000 can handle a greater density of users, but this configuration also places a lot more demand on the PCI bus, creating a possible bottleneck. Also at nearly 400 watts, excess heat could certainly be an issue, especially for densely configured servers.

Overall, AMD has added an interesting offering to its portfolio that will likely appeal to a certain class of users. However, in their zeal to go head-to-head with NVIDIA's Kepler architecture, AMD may have rushed the launch. Questions about partner OEM and software support were met with a wait-and-see response. We know OpenCL 1.2 is supported but the company was mum when it came to further parallel programming languages and tools.

As for application performance, neither real-world nor artificial benchmarks were available at press time. To be fair, AMD said that more would be revealed during SC, and they even hinted that we should pay close attention to the next Green500 list. We shall see!