Before general-purpose GPUs broke onto the supercomputing scene a few years ago, HPC was nearly a monoculture, processor-wise. According to the latest IDC numbers, x86 CPUs own about three-quarters of the market by revenue share. But the annual Supercomputing Conference always manages to showcase a number of exotic machines based on more esoteric parts, and this year’s event in New Orleans was no exception.
IBM’s Blue Gene/Q
First up is IBM’s Blue Gene/Q, the company’s next-generation architecture in the Blue Gene lineage. If you happened by the IBM booth at the conference, you could get a look at the compute and the I/O boards that will go into the upcoming BG machinery. The compute box has 32 nodes inside, with each node containing 16 PowerPC cores and each core able to manage four threads simultaneously. With 16 GB per node, IBM has managed to maintain a core:GB ratio of 1:1.
All of this adds up to a lot of flopping horsepower. According to IBM Deep Computing VP Dave Turek, just four BG/Q racks will deliver a petaflop (sustained) of performance. And true to the Blue Gene lineage, it does so with minimal power. A prototype BG/Q system topped the latest Green500 list announced at SC10 last week, with a figure of 1,684.2 megaflops per watt. That beat out even the power-sipping TSUBAME 2.0 super at Tokyo Tech, which derives most of its computational muscle from energy-efficient NVIDIA Fermi GPUs.
The first Q — at least the first big one — will be installed in 2011 at Lawrence Livermore National Laboratory. That system, known as Sequoia, is intended to be THE big production machine for the NNSA’s weapon simulation codes maintained under the Advanced Simulation and Computing (ASC) Program. Sequoia is slated to be a 20-petaflop system when it boots up next year.
Fujitsu’s K Supercomputer
Also on display at SC10 was Fujitsu’s Sparc64 VIIIfx CPU. The processor will be the basis for Kei Soku Keisanki, aka the “K computer.” That machine will be the culmination of Japan’s Next-Generation Supercomputing Project and is expected to deliver 10 petaflops for its main customer, RIKEN (Japan’s Institute of Physical and Chemical Research). Originally scheduled to come online in 2010, the pullout of NEC and Hitachi last year pushed the timeline out significantly. Full deployment is now scheduled to complete in 2012.
The Sparc64 VIIIfx processor itself is an 8-core scalar CPU that can deliver 128 peak gigaflops. Energy efficiency appears to be quite respectable. A cut-down K computer system at RIKEN was ranked number four on the latest Green500 list at 828.67 megaflops per watt (or about half that of the BG/Q prototype). Keep in mind, the original K machine was supposed to contain vector hardware as well, but when NEC and Hitachi bailed, the scalar CPUs from Fujitsu were forced to carry the entire computational load. The current design puts 80,000 Sparc64 VIIIfx chips into the 10-petaflop machine.
Tilera’s Manycore Processors
For a company that immodestly claims on its website that it “has solved the multi-processor scalability problem,” one would expect the supercomputing crowd to take notice. And it has. Tilera, makers of manycore microprocessors, has managed to attract both SGI and DARPA for HPC duty.
The 64-core Tilera processor will be an option on SGI’s Prism XL supercomputer (the offspring of Project Mojo), the company’s new accelerator-centric platform unveiled at SC10. Although most Prisms will likely be outfitted with GPGPU, SGI determined that the power-sipping Tilera silicon would be a great fit for HPC-style workloads that mainly need integer acceleration — apps like encryption, image and signal processing; network packet inspection, Web/content delivery, and media format conversion. Whether this particular configuration catches on or not remains to be seen, but you have to give SGI credit for going after market niches that other HPC vendors have largely ignored.
Tilera is also a player in DARPA’s Ubiquitous High Performance Computing (UHPC) program, where its manycore tiled processor technology garnered the company a place on one of the four initial teams. Anant Agarwal, Tilera co-founder and CTO (and EE/CS professor at MIT) pitched his UHPC team’s Project Angstrom at a Friday panel at SC10. In his presentation, Agarwal emphasized the performance per watt strength of the Tilera technology versus conventional CPUs. For example, to attain a targeted 50 mW (milliwatts) per core performance needed for UHPC machines, Tilera has only to modestly scale its current 200 mW per core designs. Agarwal proposes they can deliver that on 11nm technology with a 1,000-core 50 watt chip that delivers 5 teraflops. Conventional CPUs, he argues, are going to have to undergo a deeper architectural redesign, given that they currently consume around 10 watts per core.
Cray’s XMT supercomputer has been around since 2007, but has always been overshadowed by the company’s mainstream XT and now XE lines. Outside the three-letter intelligence agencies and a few US DOE labs, the machine is not widely known. But Cray is apparently looking to expand its popularity. At SC10 this year, the XMT got some extra attention, appearing in the Disruptive Technologies exhibit and as the focus of its own BoF.
The XMT’s forte is scalable data analytics, and the architecture has been designed with this application set in mind. Encompassing Cray’s custom SeaStar2 interconnect and Threadstorm processors, the platform’s principle architectural feature is globally-addressable memory, which makes it possible to run shared memory applications on the machine. All Threadstorm chips can access each other’s memory (up to 8 GB per processor) making it feasible to build a system with as much as 64 terabytes of global RAM.
Unlike most shared memory machines, the XMT is built to support a lot of parallelism. Each Threadstorm CPU can manage 128 threads at a time. Combined with speedy access to random chunks of remote shared memory, the system is a much more efficient platform than a conventional distributed memory architecture for those applications that require processing of really big graph-oriented databases. This includes a lot of large-scale scientific data analysis as well as many non-technical informatics applications where the data is unstructured. Look for Cray to keep pushing this technology into this rapidly emerging space.
SeaMicro’s Atom-Based Server
Perhaps the most obscure exotic at SC10 was SeaMicro’s SM10000, an Intel Atom-based server that puts eight of the tiny processors onto a card, 512 in a 10U enclosure, and 2,048 in a rack. The Z530 processor being used is a single core, 1.6 GHz chip that has a max TDP of a mere 2 watts. The company’s pitch is that this setup requires just one-quarter of the power and space of conventional x86 servers without requiring any modifications to software. Atom is conveniently x86 compatible.
The downside is that it’s a pretty low-end set-up. Memory maxes out at 2 GB per card; network support is Ethernet only; and the single-core chip in the current version is 32-bit. That may be acceptable for low-precision throughput apps that need lots of parallelism but don’t require any sort of tight coupling or single-threaded performance. A 64-bit version with InfiniBand or low-latency 10GbE connectivity would be a much more interesting offering. But keep on eye on the Atom. It could be the dark horse in the race for energy-efficient x86 HPC.