CHIP DESIGNER Nvidia has launched its Tesla K20X GPGPU accelerator card aimed at double-precision floating point scientific workloads.
Nvidia had already announced that its Tesla K20 series would be used in what was forecast to be the most powerful high performance computing (HPC) cluster, Titan, a few weeks ago. Now the firm has revealed the launch its Tesla K20X double-precision floating point GPGPU accelerator card and confirmed that the Cray XK7 Titan cluster at the US Oak Ridge National Laboratory, which boasts 18,688 Tesla K20X GPGPU cards, has taken top spot on the prestigious Top 500 HPC list.
While Nvidia's Kepler architecture made its first showing in the HPC market with the single-precision Tesla K10, the firm waited over six months before bringing out the big guns. The firm's Tesla K20X card delivers 1.31 TeraFLOPS of peak double-precision floating point computing power and 3.95 TeraFLOPS of peak single-precision floating point computing performance, and is supported by 6GB of onboard memory with a bandwidth of 250GB/s.
Nvidia also announced the Tesla K20, a marginally cut-down version of the Tesla K20X that boasts peak single-precision floating point computing power of 3.52 TeraFLOPS and 1.17 TeraFLOPS for double-precision floating point. The firm cut back memory capacity to 5GB and memory bandwidth to 208GB/s on the Tesla K20 compared to its top-end Tesla K20X.
Nvidia also took the opportunity to take a few pops at Intel's latest Xeon Phi accelerator by claiming that there is limited support in terms of libraries. However, in what might be the pot calling the kettle black, Nvidia pointed out that Intel has proprietory directives and language extensions, two things found on Nvidia's CUDA, though the firm now supports OpenCL on its GPGPU accelerator cards.
While Cray and the Oak Ridge National Laboratory got their hands on Nvidia's Tesla K20X parts months ago, the firm said that the Tesla K20 and Tesla K20X GPGPU accelerator cards are now available for general order from a number of server vendors including Asus, HP, IBM, SGI, Supermicro and Tyan. µ