CHIPMAKER Intel has for the first time in years taken first place in the prestigious Top 500 supercomputers list with the all-Intel Tianhe 2 cluster using Xeon processors and Xeon Phi accelerators.
Intel's Xeon Phi GPGPU accelerator is a descendant of its Larrabee research. Intel introduced it a year ago at the International Supercomputing show and it didn't take long for the firm to put enough of the boards into a cluster to help it reach the top spot. The firm sits on top of the Top 500 with the Chinese Tianhe-2 cluster, which is powered by Xeon E5-2692 processors and Xeon Phi 31S1P accelerators, and has a Linpack score of 33.86 petaFLOPS.
The Chinese National University of Defense Technology houses the 3.12 million core Tianhe-2 cluster, which has one petabyte of RAM, uses a proprietory interconnect dubbed the TH Express 2 and consumes 17.8 GWatt of electricity. Like most of the clusters on the Top 500 list, the Tianhe 2 runs a custom Linux distribution.
Intel already had one cluster in the top 10 of the Top 500 with Stampede, however the firm has raised the bar significantly with the Tianhe 2, which is close to double the performance of Titan, the number two system on the list. Cray's Titan cluster previously topped the list with AMD Opteron processors and Nvidia Tesla K20X accelerators.
While Intel capturing the top of the Top 500 list is an impressive feat given the relative immaturity of Xeon Phi, Cray and Nvidia can rightfully claim that the Titan cluster could also hit the same 33.86 petaFLOPS and use less power, as Titan uses 8.2 GWatt to hit 17.6 petaFLOPS. In fact, going through the full Top 500 list, AMD can lay claim to powering the most energy efficient cluster.
Adtech's ESC4000 cluster that is installed in Saudi Arabia uses Intel Xeon E5-2650 processors and AMD Firepro S10000 accelerators, and while the cluster tips up at number 52 on the Top 500, it has a MFLOPS/Watt of 2,973, over 20 percent higher than the second place Xeon Phi powered cluster, while the Nvidia powered Aurora disappointed, coming in at just 2,193 MFLOPS/Watt, considerably lower than chip designer Nvidia or cluster builder Eurotech predicted.
Intel has much to shout about with the Xeon Phi, as the firm has gone from practically a standing start in the accelerator business to having a third of the top six in the space of a year. Now if AMD would start to take its accelerator business seriously, it could push Intel and Nvidia hard, given its energy efficiency. µ