Power 7 Will Be IBM’s Supercomputing King

This weekend I heard a few guys from IBM India getting really excited about the next-generation microprocessor called Power 7 which will be released in H1 2010. The Power 7 processor is still not a final design but nonetheless, it is turning out to be an awesome processor. IBM’s latest plans show that Power 7 will be a 8-core processor, clocked at anywhere between 3.6Ghz-4.0Ghz.

Power 7 is a completely new architecture and is being developed at Austin, Böblingen, Rochester Research Centers. There are rumors that Opteron and Power 7 may share a common socket, but that doesn’t seem to be the case in the current design of the processor. But I got hints that things on the socket front could be changed during the final few weeks.

Power 7 is also a multi-thread per core CPU and current designs suggest it can execute about 4 threads per core. It still isn’t running optimally at the moment, but the target is to achieve full performance on all the 4 threads. This still makes Sun’s Rock the multi-threading king in the next-generation of processors. But IBM is high on the number crunching game. Each Power 7 core will be 32GFlops and that means on a 8-cores it going to output at 256GFlops. That’s no mean achievement!

Unlike the Power 6, which had its main aim towards improving single-core performance and max clocked at 6Ghz in prototypes, the Power 7 is working towards a multi-core, multi-threaded processor. From Intel to Sun, everyone has realized the potential efficiency of multi-core design and IBM hopes to lead the pack with Power 7. And to become the winner of the lot, Power 7 will compete against Sun’s Rock (UltraSPARC T2 Processors), Intel’s Itanium and Xeon (Nehelam-based) to some extent as well as its friendly technology partner AMD’s Opteron. IBM’s main goal with Power 7 is at the High-Performance Computing markets and supercomputers, where IBM already rules the roost.

If you look for stats at Top500 Supercomputers from around the world, you’ll realize what I mean by IBM ruling the roost. IBM has lots of BlueGene machines that use the old PowerPC processors, but not many newer generation processors. The reason to this is two-fold. IBM’s last generation processor Power 6 and to some extent Power 5 isn’t very power efficient. Power 6 heats up quite a bit and would require massive engineering skills to bring 1000s of these chips together inside a supercomputer. On the other hand, IBM has plans to mix-match Power 7 and Cell microprocessors in future supercomputers. Cell processors are known to be good specialized FP units and matching it with Power 7, IBM hopes to reach more than 10 petaflop from a supercomputer.

Intel Steals AMD's Supercomputer Shopper

Cray and Intel announced today that they have established an R&D deal for research on next generation supercomputers. Cray Inc is one of the best supercomputer and high-performance computing manufacturer and mainly has been using AMD processors in its machines. Cray and Intel will be researching on the use of Larabee accelerators in supercomputers.

Larabee is Intel's next-generation GPGPU initiative, which is the cunning plan to use graphic cards to do other computing tasks than just showing graphics. GPU are known to have high-floating point processing capabilities and the industry has decided its time to use it for some math. AMD's ATI acquisition was on similar grounds and AMD today manufactures AMD Stream Processors that do the same thing. Nvidia with its CUDA is also doing a same thing.

AMD must be pretty sad that it hasn't been able to pull something similar with its Stream Processors. AMD Opteron has been pretty popular among supercomputer manufacturers due to scalability brought through HyperTransport architecture. Cray shares seem to have risen due to some excitement with AMD's chip availability, but I suspect its more to do with the Intel partnership.

Stream Processing are going to be the future of high-performance computing. Even on desktops, it'll be either be processors doing GPU work or GPU doing processors work sometime in the future.

India's Tata Owns Asia's Fastest Supercomputer

Guest Blogger: Steven Parker
13th Nov, New York

I know a lot of the reader's at this blog are Indians and this one is for all the Indians to make you feel proud. Tata Group, which is India's popular conglomerate own the fastest supercomputer in Asia and the 4th fastest in the world. The supercomputer called "Eka" (One) is located at the Computational Research Laboratories (CRL) in Pune, a wholly owned subsidiary of Tata Sons.

The rankings were announced yesterday at the Supercomputing Conference held at Reno, NV. The list can be found here in all its glory. Its is the first time that a supercomputer from India has figured in the top 10 list of supercomputers. Eka is built by HP and uses Cluster Platform 3000 BL460c, Xeon 53xx 3GHz, Infiniband. It reaches a max of 117.9 TFlops and Rpeak of 170TFlops. It moved from its earlier ranking of 179 to rank 4, and is quite a big achievement. It cost the Tata's $30 million to build. Tata's announcement can be found here about the entire thing.

I do not agree with Sunny's view of using GPUs for supercomputing. The cost of the hardware would come down considerably, but the cost of software would be a lot higher considering the productivity of programming the GPU. Let's see how this one turns out!