AMD appears to be making good on its goals of, with the company and now dropping two new graphics cards that will provide major competition in the supercomputing space.
Dubbed the FirePro S9050 and S9150, the server GPU will compute floating-point to floating-point with Nvidia's Tesla board, which has seen for the AMD rival. While the Tesla K40, Nvidia's current top single-GPU-server performer, offers 12GB of GDDR5 RAM and 288GB/sec memory bandwidth, the S9150 includes 16GB of GDDR5 memory and 320GB/sec memory bandwidth. That helps AMD's new board outpace the K40 in both single precision floating-point performance (5.07 TFLOPS versus 4.29 TFLOPS) and double precision floating-point performance (2.53 TFLOPS versus 1.43 TFLOPS). The S9050 is no slouch, either, coming with 12GB of GDDR5 RAM and 264 GB/s memory bandwidth to provide 3.23 Tflops of single precision floating-point performance and 806 GFLOPS double precision floating-point performance, which is more in line with the Tesla K20.
AMD is touting the fact that the S9150 can deliver that performance while requiring the same maximum power (235 watts) that the K40 needs. The less-powerful S9050 sips a little less juice, with a maximum power requirement of 225 watts. Both cards' computing cores are based on the company's Graphics Core Next architecture, which is found in consumer GPUs like theand workstation boards like the . For the high-performance computations clients will require, they support OpenMP 4.0 and the forthcoming OpenCL 2.0 framework, and as Nvidia has , AMD has Stream technology to optimize the FirePro's workflow.
A number of partners have been lined up to provide servers equipped with the new FirePro GPUs, including Asus, Gigabyte, and Supermicro. They will be competing with Nvidia-based supercomputers, such as the forthcoming systems thatalong with Tesla cards. No pricing has been announced for the S9050 or S9150, but they are expected in the third quarter of this year.