The Week in Review
Here is a collection of highlights from this week’s news stream as reported by HPCwire.
Tokyo Tech heralds Tsubame 2.0′s debut
The supercomputing world better look out, Japan is about to take a big seat at the petascale table. An article from Asian tech pub Tech-On relays the latest details regarding the Tsubame 2.0 supercomputer, which the Tokyo Institute of Technology released this week during a press meeting. According to the university, the system will have 2,816 six-core, 2.93 GHz, Intel Xeon 5600 microprocessors and 4,224 NVIDIA Tesla M2050 graphics processing units (GPUs). Performance per node comes out to 1.6 teraflops and performance per rack is 51.2 teraflops. The system will be constructed by NEC and Hewlett-Packard and will cost approximately $35 million, including basic maintenance for four years.
Such impressive hardware specs place the Tsubame 2.0 securely in the land of petascale. In fact, Japan’s first petaflop system prepares to easily leap past the one petaflop mark to clock in at 2.39 petaflops, which could earn the supercomputer a second finish on the TOP500 list, but alas, that’s only in terms of theoretical peak performance. As of the June 2010 list, the highest-rated systems are China’s Nebulae, which clocks in at 2.98 petaflops peak, and the United State’s Jaguar, which clocks in at 2.33 petaflops peak. Of course, the primary TOP500 measure is maximum Linpack performance, by which the Tsubame 2.0 system will likely still achieve a very respectable third of fourth place finish. However, since the TOP500 is a moving target, almost anything is possible.
Satoshi Matsuoka, a professor at the Global Scientific Information and Computing Center (GSIC) at Tokyo Tech, explains that while the system has a mixed vector-scalar architecture, it actually functions more like a vector computer because the computation capacity of its GPUs accounts for 90 percent of total computation capacity. Because of this architecture, the computation capacity will differ depending on the type of calculation it must perform. In terms of the Linpack benchmark, optimum performance is between 1 and 1.4 petaflops (double-precision value). However, for vector computer type calculations, such as weather prediction, the performance can reach the 150 teraflops range, vastly outperforming the the world record of 50 teraflops.
In comparison with Tsubame 1.0, which was constructed in 2006, the second-generation supercomputer touts improved memory bandwidth at 200 Tbps, 33 times higher than its predecessor. Total memory bandwidth is 720 Tbps, a 42x improvement over the Tsubame 1.0. The supercomputer’s multilevel storage system is comprised of DDR3 DRAMs and SSDs (solid state drives). The memory capacity of the backbone system’s DRAMs is 80.6 terabytes for the microprocessors and 12.7 terabytes for the GPUs, and the total memory capacity of the SSDs is 173.9 terabytes.
The new system also touts improved power efficiency versus its predecessor, due in large part to the GPUs decreased energy needs, as well as the machine’s sealed cooling system. The power consumption of the Tsubame 2.0 is 1 MW versus its predecessor’s energy requirement of .85 MW; however, considering the newer supercomputer has a performance that is 30 times greater than that of its predecessor, the overall power consumption is lowered to 1/25th, a huge savings. If performance per watt meets expectations by exceeding 1,000 megaflops per watt, the supercomputer will reserve a possible top spot on the Green 500 list, according to Tokyo Tech officials.
Despite the fact that Tsubame 2.0 is expected to start production this fall, the actual construction of the system has yet to commence. Meanwhile, plans are already underway for Tsubame 3.0, with a projected debut of 2014 or 2015, a proposed 30 petaflops of computing muscle, and an energy requirement that is equal to or less than that of the second-generation Tsubame.
Cloud-optimized servers announced
There has been a spate of so-called cloud-optimized servers announced in the last couple of weeks. Things started off last week with newcomer SeaMicro’s introduction of its Internet-optimized server, the SM1000, which makes use of low-power Atom processors to handle Web-centric workloads. This week, Tilera and Quanta unveil their own server comprised of many-core general purpose microprocessors aimed at tackling cloud computing workloads. The Tilera/Quanta server, codenamed S2Q, is targeted at large-scale datacenters running Web, database, hosting, and finance applications.
Like SeaMicro’s “Internet-optimized” x86 server, Tilera’s many-core design works with cloud applications, which execute millions of small parallel tasks simultaneously, instead of very complicated single-threaded programs that require very big cores. Each server includes eight Tilera TILEPro64 processors, and, according to Tilera, is the most power efficient and highest compute density 2U server in the industry.
Last up in the run on cloud server announcements is AMD’s new Opteron 4000 Series platform, which AMD says is the “first server platform designed from the beginning to meet the specific requirements of cloud, hyperscale datacenter, and SMB customers needing highly flexible, reliable, and power-efficient 1 and 2P systems.”
These new architectures all aim to address the divide between today’s servers and the current dominant workload, which has gone from a few very difficult tasks to lots of smaller tasks as required by today’s Internet (or cloud) applications that we are all so familiar with (searching, email, social networking, etc.). For more on this subject, see related coverage at our sister publication HPC in the Cloud, with Editor Nicole Hemsoth at the helm. Here are some of her thoughts on the topic:
Granted, while [the Tilera servers] don’t have x86 compatibility, if they are able to demonstrate solid power and performance this will certainly be attractive for many, especially as large-scale datacenters are having trouble fitting their current servers in at the rate they’d like to. Intel and AMD are making the same play, but they are certainly lagging behind Tilera’s much more aggressive manycore roadmap. It is worth noting as well that SeaMicro who emerged from the ether last week with its announcement of a low-power server sporting 512 Intel Atom cores is another player here, as are those working on ARM chips for similar types of servers. In the case of SeaMicro, however, while they are functioning on the same general idea, they will likely not have the same power-performance zing since they will be tied inextricably to the legacy x86 architecture.