AMD Unveils Fast GPU Server Accelerator for HPC

Advanced Micro Devices this week unveiled a new powerful graphics card aimed at servers running in high-performance computing environments and that will compete with Nvidia’s Tesla K40 GPU.

According to AMD officials, the FirePro S9150 server card offers up to 2.53 teraflops (one trillion floating point operations per second) of double-precision peak performance aimed at highly compute-intensive and parallel scientific workloads. By contrast, Nvidia’s Tesla K40 hits a peak performance of 1.43 teraflops.

The single-precision peak performance of the AMD GPU is 5.07 teraflops, compared with 4.29 teraflops for Nvidia’s offering.

Research institutions and other high-performance computing (HPC) increasingly are embracing high-end systems that leverage GPU accelerators and co-processors to enhance the performance of those systems while keeping the power consumption in check. Of systems on the June Top500 list of the world’s fastest supercomputers, 62 use accelerators or co-processors, an increase from 53 in November 2013. Forty-four of those systems included Nvidia GPUs, while two used Radeon chips from AMD. Another 17 leverage x86-based Xeon Phi co-processors from Intel.

Four of the top 10 systems on the list used either GPU accelerators from Nvidia or Xeon Phi chips from Intel. The fastest system in the world, the massive Tiahne-2 system at the National Supercomputer Center in China, included Xeon Phi co-processors, while the second-fastest—the Cray Titan system at the Oak Ridge National Laboratory in Tennessee—is powered by AMD’s x86-based Opteron processors and Tesla K20 GPU accelerators.

“Today’s supercomputers feature an increasing mix of GPUs, CPUs and co-processors to achieve great performance, and many of them are being implemented in an environmentally responsible manner to help reduce power and water consumption,” David Cummings, senior director and general manager of professional graphics at AMD, said in a statement.

AMD built the FirePro S9150 to offer high performance, more memory and increased performance per watt, Cummings said.

Other features of AMD’s FirePro S9150 GPU include 16GB of GDDR5 memory, 512-bit memory interface and up to 320GB per second of bandwidth, according to company officials. It’s based on AMD’s Graphics Core Next (GCN) architecture, can support the upcoming OpenCL 2.0 parallel-processing framework and has a maximum power consumption of 235 watts.

Another version of the chip, the FirePro S9050, offers 12GB of GDDR5 memory, a 384-bit interface and 264G bps of bandwidth. It also is based on the GCN architecture and supports OpenCL 1.2. It brings 3.23 teraflops of single peak precision performance and 806 gigaflops of double-precision performance while consuming a maximum of 225 watts.

The University of Bristol’s Microelectronics Research Group has been testing the FirePro S9150, and Simon McIntosh-Smith, head of the Microelectronics research group, in a statement called it “the fastest HPC GPU accelerator we’ve ever tested” while lauding AMD’s decision to include support for OpenCL 2.0, which will come later this year.

Vendors committed to using the GPU server card include Asus, Gigabyte and MiTac International’s Tyan business unit. In addition, Supermicro is planning to leverage both the S9150 and S9050 in its Green Computing portfolio of HPC systems.