2.5 PFLOPS: China Deploys World's Most Powerful Supercomputer

Tianjin is becoming a modern city for the 21st century. This 12-million inhabitant city is currently building a Maglev-based Metro for its inhabitants and is becoming a center of supercomputing as well. Scientists at NUDT in Tianjin deployed Tianhe-1A, the world's most powerful supercomputer based on nVidia Tesla GPGPU and Intel CPUs.

The system is based upon 2:1 ratio between CPU and the GPU, meaning that 14,336 Intel processors feed 7,168 Tesla M2050 GPGPU cards. The order for Tesla cards alone was worth $17.92 million at list price. The scientists at the university did not reach for world's fastest CPUs, meaning the University deployed more affordable Xeon processors, meaning less revenue for other parties involved.

This combination results in 2.51 PFLOPS of sheer computing performance measured by the industry standard LINPACK benchmark. National University of Defense Technology did not disclose the theoretical peak, as that information will wait until the unveiling of November Top500 list.

According to Top500 list from June 2010 [new Top500 list will be revealed on Nov 13 at SC'10 in New Orleans, LA], Tianhe-1A sits comfortably at the top of the list as the world's most powerful supercomputer to date.
This is not the only GPGPU-powered supercomputer coming from National University of Defense Technology from Tianjin. If you recall, in 2009 NUDT deployed Tianhe-1 which combined older 45nm Xeons with now-AMD Radeon HD 4870 X2, dual-GPU graphics cards. Tianhe-1 reached 563 TFLOPS and served as a warm-up for a complex task of making 7,168 GPU cards work in sync with 14,336 processors.

Timing is everything, and so is today's announcement. Tianhe-1A also marks the first time in history that US might lose the #1 position in Top500, given that Jaguar from Oak Ridge National Laboratories will probably remain at 1.76 PFLOPS [Rmax] and even with a theoretical peak at 2.33 PFLOPS, it shouldn't be able to keep the lead over Tianhe-1A. If you're an American and worried about this shift, you'll probably be happy when you hear that Oak Ridge is currently setting up a system based on more than 15,000 Tesla cards, targeting the 10 PFLOPS. In order to break that barrier by using Tesla M2050 cards, Oak Ridge needs at least 28,672 cards worth 71.68 million USD at list price.

The trend behind GPGPU i.e. General Purpose GPU computing is shaping up as an avalanche, rather than just an appearing trend. Bear in mind that at the back end of 1990s, x86 architecture took over the proprietary-laden world of HPC [High Performance Computing] by storm, headed by Pentium II and Pentium III based Intel Xeons. In only 18 months, x86 captured 80% of the market and essentially ended the proprietary CPUs. Today, even mainstream console sales cannot help the volume of Power architecture when compared to the volume reached by CPUs from Intel and AMD, as well as GPUs coming from nVidia and AMD respectively.

Today, it looks like GPU computing is headed in the same direction, as GPUs now power two out of three world's most powerful supercomputers [world's #3 is occupied by Shenzen-based Nebulae, 1.27 PFLOPS].