by Sumit Gupta

Tesla GPUs are already used worldwide to boost supercomputers, and now they’re being used to crank up performance on new servers based on Intel Sandy Bridge CPUs.

We just got our hands on the HP SL250 ProLiant server, and we took it for a test drive running some popular scientific high-performance computing (HPC) applications. As expected, the Tesla GPUs significantly accelerated the performance of the system.

The HP server includes 3 Tesla M2090s GPUs, as well as new Intel Sandy Bridge CPUs. The new CPUs provide higher performance than older versions when it comes to processing sequential (serial) tasks. This frees up the GPUs to attack and complete parallel tasks significantly faster.

Here are some early benchmark results, using the popular molecular dynamics applications NAMD and LAMMPS:

Some highlights:

  • Both NAMD and LAMMPS scale almost linearly from 1 GPU to 3 GPUs in the server.
  • Adding just one GPU to the server leads to an almost 2x speedup across all the benchmarks.
  • Three GPUs provide speedups of at least 5x compared with using two CPUs alone for all the benchmarks. Three Tesla M2090s GPUs accelerated the Gay-Berne (GB) method in LAMMPS by nearly 21 times over CPUs.

This is the configuration of the HP SL250 ProLiant server we tested:

  • Dual socket Intel Sandy Bridge (2.2 Ghz) CPUs (total 16 CPU cores)
  • Two Tesla M2090 GPUs
  • 66 GB DDR3 ECC
  • Redhat Enterprise Linux 6.1
  • NVIDIA driver version 295.20

If you have a Sandy Bridge server with Tesla GPUs, I encourage you to run benchmarks on your own.

Here is the data used to create the chart above:

Running your own benchmarks on your own Sandy Bridge + Tesla GPU system? Leave a comment below with your performance results.