The new edition of TOP500 is out – and it is a surprise to see AMD heavily featured in the fastest supercomputers. Of the top 5 supercomputers in the world – AMD products find there way into all but one. Surprisingly, we have to look at #5 for the first showing of Intel.

Architecturally, Intel’s Nehalem based Xeon CPUs are vastly superior to AMD’s Shanghai based Opterons. However, six core Opterons do have a core count advantage. Even with 50% more cores, the quad core Xeons often sneaks past their Opteron rivals. There are applications which prefer the higher core count – and perhaps this is something supercomputer makers consider.

More next page.

The new edition of TOP500 is out – and it is a surprise to see AMD heavily featured in the fastest supercomputers. Of the top 5 supercomputers in the world – AMD products find there way into all but one. Surprisingly, we have to look at #5 for the first showing of Intel.

Architecturally, Intel’s Nehalem based Xeon CPUs are vastly superior to AMD’s Shanghai based Opterons. However, six core Opterons do have a core count advantage. Even with 50% more cores, the quad core Xeons often sneaks past their Opteron rivals. There are applications which prefer the higher core count – and perhaps this is something supercomputer makers consider.

The world’s fastest supercomputer is the Jaguar, previously powered by quad core Opterons – now featuring over 37,000 six core Opterons. The result is an incredible 2.33 PFlops theoretical performance, and 1.73 PFlops through the LINPACK benchmark. Previous champion Roadrunner is pushed down to second, featuring a mix of Cell and Opteron CPUs. AMD make it three in a row with Kraken – which is a basically Jaguar lite with fewer six core Opteron CPUs. In #4 is JUGENE which features neither Intel or AMD – but IBM’s Power processors.

Finally, at #5 we have China’s new Milky Way One supercomputer. This also marks the first Intel appearance, powered by Xeon CPUs. However, AMD is present once again – this time in GPGPU form, with over 2500 ATI Radeon HD 4870 X2s. Just 2500 of these GPUs offer an enormous theoretical compute power – third only to Jaguar or Roadrunner. However, as has been the case, harnessing the power of these parallel processors has been a challenge. The Milky Way One’s benchmark performance ends up less than half of the theoretical peak. (as opposed to ~80% for CPU based supercomputers).

With better documentation and development tools for OpenCL, DirectCompute 11 and CUDA alike, we can expect the fastest supercomputers of tomorrow to gradually incorporate GPGPUs. Certainly, this is something Nvidia is counting on, as Fermi will make a potent processor for a supercomputer.

Reference: Techreport

Update: While AMD has captured the very fastest supercomputers, Intel claim their processors are present in 80% of the Top 500. Not unexpected – though it was certainly a surprise to see significant AMD presence at the top end of the list.