CHIP DESIGNER Nvidia's Tesla K20 GPGPU card has tipped up in the Titan supercomputer cluster based on Cray's XK7.
Cray's XK7 is the latest model of the firm's X series of high performance computing (HPC) systems, and in the Titan supercomputer it will replace the Jaguar cluster at the US Department of Energy's Oak Ridge National Laboratory (ORNL). Titan is also the first outing for Nvidia's Tesla K20 GPGPU cards. These are mounted in 18,688 nodes, each comprising one 16-core Opteron 6200 series CPU paired with a Tesla K20 card, all wired together with Cray's Gemini interconnect to deliver peak computing power of roughly 20 PetaFLOPS.
ORNL's Titan cluster is the first Cray XK7 based supercomputer cluster and for Nvidia it is the first outing of its double-precision floating point Tesla K20 GPGPU card. Nvidia announced the single-precision floating point Tesla K10 card earlier this year and with the Tesla K20 the firm now boasts an impressive high profile deployment, albeit one that has yet to be fully accepted by its customer.
For ORNL, the Titan cluster will represent a significant increase in performance, going from the 2.3 PetaFLOPS provided by Jaguar to an estimated peak of 20 PetaFLOPS on Titan. However Nvidia wouldn't say what the sustained performance of Titan is and Cray was careful to point out that the Titan cluster has not been accepted yet by ORNL and wouldn't forecast when that will occur.
While ORNL's Titan cluster contains 299,008 Opteron cores, Sumit Gupta, GM of Nvidia's Tesla Accelerated Computing business unit said the vast majority of computational power comes from the Tesla K20 boards. "90 percent of the 20 PetaFLOPS is coming from the GPU, so the CPU is really playing a supporting cast here. Most of the performance is coming from the GPU," claimed Gupta.
It isn't known whether ORNL's Titan cluster will take the top spot in the upcoming Top500 list to be announced at the International Supercomputing show in November, but given that Sequoia, which currently holds the top spot, has a peak computational performance of 20.1 PetaFLOPS, it will be a close contest. However Gupta claimed Titan will be the top cluster that will enable "open science" research, meaning that universities will be able to apply for access to the cluster, something that many high-end government sponsored HPC clusters do not allow. µ