Ever since the HP ProLiant SL powered TSUBAME2.0 system debuted at #4 on the November 2010 Top500 list, the ProLiant SL platform has been one of the world’s leading purpose-built servers for GPU computing. Since Nvdia’s latest Kepler line of GPUs has been shipping for several months now, I thought it would be good to give an update on how HP customers are adopting Kepler.
HP currently offers three different Kepler GPUs, the K10, K20, and K20X. These GPUs are available in HP’s ProLiant SL250s and ProLiant SL270s servers. First, it is important to understand Nvidia’s naming convention. The K10 is not an entry level into the Kepler line but essentially a special purpose GPU optimized for certain types of applications requiring primarily single precision (vs double precision) performance. One of the largest adopters of the K10 has been the oil & gas industry, where one of the primary workloads, Reverse Time Migration, or RTM, is dominated by single precision math. So it should come as no surprise that the K10 has been a great seller with our worldwide oil and gas customers.
The K20, on the other hand, is Nvidia’s mainstream GPU and if you have used earlier Nvidia GPUs like the M2070 or M2090, you will love the K20. HP has shipped several mid-size (by server count) GPU clusters including a 264 node SL250s system to USC with 528 K20 GPUs and a cluster of over 100 SL250s servers each also with two K20 GPUs to Clemson. The K20X GPU is similar to the K20, only with more memory and more GPU cores and we currently have multiple customers testing K20X systems.
Before we shipped USC their 264 node cluster, HP’s High Performance Computing benchmarking team had a few days to use the system in our factory to get a couple of Linpack runs completed. While I won’t share the exact numbers, the system would easily rank in the Top50 on the current Top500 list. That is a pretty amazing testament to the power of Nvidia’s Kepler technology when with only 264 servers you can build one of the 50 fastest supercomputers in the world (given the semi-annual performance improvements in the Top500 list, I won’t make any projections as to where the USC system will land in the next Top500 list to be published this coming June). Compared to earlier generations of GPUs, we have seen significant efficiency improvements with Kepler (measured performance / peak theoretical performance) not only on Linpack but also on a wide range of real world applications.
Expect many more updates on Kepler at Nvidia’s upcoming GPU Technology Conference, March 18-21 in San Jose, California.