Advanced Micro Devices has stepped up its supercomputer battle with Nvidia, claiming the graphics performance crown with its fastest server GPU offering yet, the FirePro S9150.
Graphics processors are at the center of performance advances in supercomputers. Working with CPUs, GPUs speed up math calculations and scientific simulations. The world’s second fastest supercomputer, Titan, uses 18,688 Tesla K20X GPUs from Nvidia.
AMD says its S9150 GPU is faster than Nvidia’s top supercomputing GPU. The S9150 offers double-precision peak performance of 2.53 teraflops, while Nvidia’s top-end Tesla K40 clocks in at 1.43 teraflops. The single-precision peak performance of FirePro S9150 is 5.07 teraflops, slightly higher than 4.29 teraflops for Nvidia’s Tesla K40.
Double-precision performance is more important for supercomputing as it represents a more accurate floating-point calculation.
AMD’s server GPU performance lead could however be short-lived. Nvidia in the coming months is expected to upgrade Tesla GPU products to the new Maxwell graphics architecture, which should push peak performance even higher than what’s offered by the K40.
AMD’s FirePro products are used in a handful of server products from companies like Dell, and has not yet made its mark on supercomputers. According to a list of the fastest 500 supercomputers in the world compiled by Top500, five systems in the top 20 use Nvidia GPUs, while AMD’s FirePro makes its first appearance in the 70th-ranked Sanam supercomputer in Saudi Arabia.
Supercomputer GPUs are not as fast as the top graphics chips used in desktops. High-performance servers usually carry many GPUs, and the characteristics of hardware need to fit within performance and power consumption constraints of data centers. Desktop GPUs can be overclocked and cooled down with liquid cooling, which is not as easy on servers.
The S9150 has 2,816 processing cores, draws 235 watts of power, and packs 16GB of GDDR5 memory. The cores are based on the same graphics technology—Graphics Core Next—used in Microsoft’s Xbox One and Sony’s PlayStation 4.
Math and scientific programs can be coded for parallel execution across CPUs and GPUs through open-source OpenCL parallel programming tools. Nvidia supports OpenCL, and offers its own CUDA parallel programming framework.
AMD also announced the FirePro S9050, a slower GPU from the same product family. It has 1,792 processor cores, 12GB of GDDR5 memory and draws 225 watts of power.
The GPUs will ship this quarter and be available through distribution partner Sapphire Technologies. Server makers Supermicro, Asustek, Tyan will offer the S9150 GPUs in products.