AMD unveils FirePro S9170 GPU for data centres with over 5.2 teraflops of performance

AMD claims leading performance for its accelerators aimed at HPC and virtual workstation applications

AMD has extended its FirePro S-Series of GPU accelerator cards aimed at servers with a new model that doubles the amount of onboard memory available for processing workloads and pushes up floating point performance to about 5.24 teraflops (Tflops).

The FirePro S9170 is due to ship to system vendors later in this quarter, and is based on AMD's Hawaii Graphics Core Next GPU architecture. AMD claims it to be the fastest single GPU accelerator card for servers on the market today.

Unlike AMD's standard FirePro adapters for workstations, the S-Series is aimed at server workloads such as high-performance computing (HPC), virtual desktop infrastructure and virtual workstation deployments in the cloud.

The 32GB of onboard GDDR5 RAM gives the FirePro S9170 double the available memory space for workloads as its predecessor, the FirePro S9150, and about three times that of its nearest competitors in the GPU market, AMD claimed.

The new card also has the edge in raw performance and energy efficiency, according to the firm.

"What makes AMD's GPUs particularly well suited for HPC boils down to four things: very high compute performance; leadership in performance-per-watt; the largest available working memory size; and our embracing of open standards, including the most recent revision of OpenCL," said Niles Burbank, senior product manager for professional graphics at AMD.

The large onboard memory is picked out as a key differentiator as it allows researchers to apply bigger problems to the GPU with less need to copy data from main memory and back again, Burbank added.

However, with 44 compute units (2,816 stream processors), the FirePro S9170 is also capable of 5.24Tflops peak single-precision floating point performance and 2.62Tflops peak double-precision floating point performance.

AMD claimed that this represents up to 40 percent greater performance for specific workloads than Nvidia's rival Tesla K80 adapter.

The FirePro S9170 is delivered as a dual-slot PCI Express card with a thermal design power of 275W, enabling the new GPU to be cooled passively in the vast majority of server platforms.

In situations where this may be a problem, the card can be limited to 235W for a small drop in peak processing performance, AMD said.

AMD regards the OpenCL framework as the primary tool for programming its FirePro portfolio, but said that it recognises that some customers have existing investments in other parallel APIs.

For this reason, AMD has partnered with developer PathScale to add FirePro support to its OpenMP and OpenACC compilers.

AMD said it has already seen success with existing FirePro S-Series cards, citing a new supercomputing cluster using its hardware at the University of Warsaw, and the Lattice-CSC cluster at the University of Frankfurt, listed on the Green 500 list as the most energy efficient supercomputer in operation.

AMD declined to specify exact prices for the FirePro S9170, saying that it will depend on channel partners.

But Burbank indicated that a price tag of somewhere in the region of $3,000 to $4,000 is typical for a GPU accelerator in this class.