Crossing the HPC Chasm
|Die shot of silicon chip included in the ‘Knights Ferry’ Many Integrated Core (MIC) architecture development platform. Intel has announced plans to deliver a product, codenamed “Knights Corner,” that will scale to more than 50 processing cores on a single chip.
General purpose graphics processing units (GPGPUs) are no longer novelties in high performance computing (HPC). Systems vendors are lining up to offer GPGPUs as complements to x86 or other base processors.
Venerable Cray, which boasts the world’s fastest supercomputer, just announced plans for GPGPUs. China, a rising HPC power, used GPGPUs to propel its biggest supercomputer into second place in the world rankings.
NVIDIA is leading the charge, but some buyers like what AMD has in store, and there’s interest in Intel’s planned Knights Corner 50-core chip, which will combine x86 cores with GPGPU-like functions from the scaled-back Larrabee project.
GPGPU programming is still not facile, and moving across the PCI Express link adds time. But GPGPUs can ignite the right subset of applications, and they can offer compelling price/performance and density. It looks as if they’re here to stay.
Steve Conway is IDC Research VP, HPC. He may be reached at [email protected].