Nvidia Gives Servers A Boost With Tesla M2090 GPU
One such server, the HP ProLiant SL390 G7, can fit up to 8 Tesla M2090s in its 4U chassis, along with 2 CPUs. Sumit Gupta, senior product manager for Nvidia's Tesla group, said in an interview that this density will help data centers make better use of space while also increasing efficiency and cutting power consumption.
Gupta also said this is the highest GPU to CPU ratio in any server on the market and a sign of things to come. "The balance of performance and power is now being dedicated more to GPU than CPU," he said. "OEMs are dedicating more time to GPU to CPU, and this is a reflection of where the market is moving."
The Fermi-based Tesla M2090 comes with 512 CUDA cores, delivers 665 gigaflops of peak double precision performance and enables 10x the application acceleration capability of typical CPUs, Gupta said. The M2090's memory bandwidth is 178 Gbps, which is 20 to 30 percent faster than its predecessor, the M2070. Typical CPUs can deliver memory bandwidth of 25 to 50 Gbps, Gupta said.
In the biotechnology field, researchers typically use computer simulations to model the labor-intensive work they would do in a lab. One commonly used application in this space is Amber, which helps researchers see how atoms inside molecules interact with each other.
Amber is useful from a research standpoint, but running Amber simulations requires a supercomputer of the type found in Oak Ridge National Labs, a CPU-only version that can deliver 46 nanoseconds of simulation per day. Using four Tesla M2090 GPUs coupled with four CPUs, Nvidia delivered performance of 69 nanoseconds of simulation per day, Gupta said.
"This is a way to democratize science, enabling researchers around the world to have access to supercomputing capability that was previously limited to few people," Gupta said.
Nvidia didn't offer pricing information for the M2090, but the M2070 is currently selling on Google Products in a range between $3,000 and $4,500.