HPCwire article on Tsubame2 rankings

http://www.hpcwire.com/hpcwire/2012-02-02/number_crunching,_data_crunching_and_energy_efficiency:_the_hpc_hat_trick.html

Intel MIC to power 10 PF super at TACC

Although the Intel MIC platform (Knights Ferry and co) will not be ready until the end of next year, it seems that Intel already scoared a big goal having been selected to power an upcoming 10PF supercomputer at TACC. The full system will provide 2 petaflops of Sandy Bridge-EP processors and 8 petaflops of Intel's Many Integrated Core coprocessors. The MIC coprocessors will apparently be preproduction Knights Corner chips. The system also means the first searious blow to NVIDIA/AMD in the accelerated computing sector. Intel's MIC chips are based on the Intel64 instruction set and are compatible with popular programming models such as OpenMP which provides a much easier transition path for legacy codes. It seems that these first generation chips will include above 50 cores connected by a ring and with high bandwidth external memory (note Larrabee legacy here). The GPGPU community has so far not taken the Intel MIC line as a serious threat, mainly (I guess) because the chips are entering this market segment quite late. The CUDA programming model enforces programming methodologies that are very interesting for particularly regular applications (i.e., those that can efficiently exploit the SM shared memories). MIC may not have the computing density to beat NVIDIA on these apps, but it may well surpass the performance on irregular apps such as those dealing with unstructured grids, graphs or other sparse and pointer-based data structures thanks to its large L2 and less strict programming model. And in my view such apps will play an increasingly important role in the future. It will be an interesting battle to watch.

More information on TACC's Stampede supercomputer can be found here.

IBM cancels Blue Waters

IBM and NCSA have issued a joint statement announcing the cancellation of the Blue Waters supercomp. I confess this has caught me as a surprise -particularly after all the publicity that this project had gathered- to the point that I thought it might be some sort of bad joke when I first saw it. But it wasn't. Now NCSA will attempt to develop a new proposal for a supercomp to be operative before the end of 2012 using the same budget. From speculation on HPCwire it seems IBM is simply no longer interested in spending so much into technology from which it does not profit, even if this means less bragging rights for them. Under this scenario I wonder about the costs that Fujitsu spent on K. K and Blue Waters have some similarities (both are non-GPU machines using chips specially well suited for HPC) and both Fujitsu and IBM might have been involved in similar costs and trade-offs.

On the other hand, the planned 20PF BlueGene/Q "Sequoia" machine, also from IBM, is still on the roadmap to be deployed. Maybe this move only means that IBM wants to focus on its BG/Q technology for future high performance computing

IBM BG/Q tops Green500

The new Green500 list, the one that ranks Top500 supercomps by their FLOPS/Watt efficiency is out, and the new #1 is again an IBM BG/Q prototype. The machine, scoring at 2,097 MFLOPS/W is about 25% more efficient than the previous BG/Q prototype, which scored 1,684 MFLOPS/W. The BG/Q design is based on low-power PowerPC A2 chips, each one containing 16 cores. Unfortunately I am unable to find more information that would explain why this new prototype is more than 2x more efficient than production GPU supercomps such as GSIC's Tsubame 2.0 (958.35 MFLOPS/W), or, even more intriguing: what did they change in between the two prototypes that boosted the efficiency in 25%?

Interestingly, Nagasaki University scored #3 deploying a GPU cluster that achieves 1375.88 MFLOPS/W. This system is based on ATI/AMD cards. Nvidia GPUs have a much higher presence in HPC, but when it comes to Green500 the gap is much closer, with two AMD systems and 3 NVIDIA systems in the top 10.

Also interesting is the fact that the top supercomputer in the Top500 list (The Fujitsu-based K Computer) is #6 in the Green500 list at a respectable 824.56 MFLOPS/W. This demonstrates the efficiency of the Fujitsu SPARC64 VIIIfx chips and the Tofu interconnect.

More information: Top500 and Green500

Riken 'K' Supercomputer No 1 in June 2011 Top500

I was greatly surprised by today's news that that Japan's K Supercomputer (the K coming from the chinese character 京 which means 10^16) is the new No 1 machine in the June 2011 Top500 list. I was under the impression that this machine was scheduled to start operating in 2012. But maybe more surprising than this early appearance are two factors: First, its impressive Peak Performance at 8.77 PetaFlops and second, its Linpack rating of 8.16 PF, which translates to a 93% Linpack efficiency! Being used to much lower rates this efficiency is extremely impressive: around 75% efficiency for CPUs and about 50% for GPUs. The 8.16PF rating means that K is more powerful than the next 5 Top500 systems combined.

Interesting is also to remember that this machine was about to be scratched last year by the Japanese government, a move that was only reversed after the Japanese scientific community angrily opposed it.

New Cray machine can scale to a peak of 50 PF

The new Cray XK6 machine, providing up to 70+ TF per cabinet when fitted with M2090 Teslas, can theoretically provide a peak 50 Petaflops with 714 cabinets. The problem? It will still take 38 Megawatts to operate it at max configuration :-)

Source info here

Quantum Computing getting closer?

There was this bit on HPCwire this week about D-Wave Systems which claims to have sold the first quantum computer to the Lockheed Martin corporation. Apparently the system includes a 128 qubit chip. Here's some information. There's also a link to a Nature paper that I assume explains in more detail how the system works.