There’s no denying that NVIDIA dominates the HPC market with their "Tesla" GPGPU cards. Even though AMD offered FireStream GPGPU cards in the past, they were never available in real volume. At the same time, we were hearing from partners that Tesla (especially server-grade Mxxxx-Series) was outselling AMD’s cards from 20-to-1 to as high as 300-to-1.
In order to address the growing compute, as well as virtual desktop infrastructure (VDI) market, AMD finally committed to launching the appropriate server-grade cards in volume. FireStream branding is now all but dead, as the new server parts are called FirePro S7000 and S9000.
Left to right: AMD FirePro S9000 6GB and FirePro S7000 4GB
The S7000 is based on a Pitcairn GPU paired with 4GB of GDDR5 memory (ECC). As you might know, Pitcairn carries a 256-bit memory interface, which will result with a compute/video memory bandwidth of 154GB/s. In terms of performance, AMD Claims 2.4 TFLOPS in Single Precision and lowly 152 GFLOPS in Double-Precision. Naturally, we’re talking about peak performance figures. This board is a single-slot design, created to maximize the airflow coming from inside the server. Power consumption? You’re looking at a 150 Watts (PCIe x16 gives 75W, 6-pin connector additional 75W). If you are interested in the part, you will have to spend $1,249.
The S9000 is based on Tahiti XT GPU paired with a 6GB of GDDR5 memory (ECC). With the 384-bit memory interface, this part will go head to head not just against the 256-bit Kepler-based Tesla M2070/M2075/M2090 cards, but against the upcoming 384-bit GK110 (Tesla K20) as well. The board has 264GB/s of compute/video memory bandwidth and gives out 3.23 TFLOPS Single-Precision, as well as 806 GFLOPS Double-Precision. If these numbers hold true, S9000 will narrowly edge out Tesla K20 in single-precision (by around 230 GFLOPS) and dramatically loose over K20 in double precision (by 693 GFLOPS). As expected from a high-end product, the S9000 is a dual-slot card and takes 225 Watts through two 6-pin connectors and the PCIe interface. In case you are interested, the price for this board is $2,499.
While it is hard to believe that AMD can achieve remarkable sales results with this first true generation server part (it took NVIDIA two generations to be taken seriously), the fact of the matter is that both S7000 and S9000 parts are cash-cows for the company. In our discussions with Investor Relations and our clients, it is clear that NVIDIA enjoys the biggest margin on the Tesla parts, especially server ones.
After all, the board cost is the lowest of the professional parts (you don’t even use the complete heatsink, no HDMI ports, single DisplayLink port), and the margin is the highest. If we take a look at competition, the current high-end M2090 6GB will set you back for $2,459, while the upcoming Tesla K20 is priced at $3,199. Naturally, the physical cost of the card pales in comparison to the human cost in thousands of man hours needed to qualify the parts. The cards are certified by Citrix, Microsoft and VMware.