October 29, 2009

China Joins Petaflop Club

Michael Feldman

China has apparently become the third country to build a petaflop supercomputer. Xinhua news agency reported on Thursday that the country has unveiled “Tianhe,” a 1.206 peak petaflop machine powered by a combination of 6,144 Intel CPUs and 5,120 AMD GPUs. Amazingly, the price tag was a mere $88.24 million. The system is installed at the National University of Defense Technology (NUDT) in Changsha, capital of central China’s Hunan Province.

In the TOP500 sense, Tianhe would not be considered a true petaflop system. According to online reports, the machine achieves only(!) 563.1 teraflops with Linpack. If that number holds up, it would almost certainly earn Tianhe a spot in the top 10 of the upcoming TOP500 list. Today there are only three systems that break the 500 teraflop barrier on Linpack: Roadrunner at Los Alamos National Laboratory, Jaguar at Oak Ridge National Laboratory, and JUGENE at Jülich Supercomputing Center. China’s top system on the current list is Dawning’s “Magic Cube” supercomputer, located at the Shanghai Supercomputer Center. With a Linpack rating of 180.6 teraflops, the Dawning machine sits at number 15.

At some future date, NUDT is slated to add “hundreds or thousands of China-made CPUs to the machine, and improve its Linpack performance to over 800 teraflops,” according to Zhou Xingming, an academician in the Chinese Academy of Sciences and a professor at NUDT.

The Xinhua news article, as well as other early reports from Chinese sources, don’t provide much detail about the system’s architecture. Specifically, no information was offered about the kind of Intel CPU and AMD GPU parts used, nor about the Chinese-made CPUs to be plugged in later on. At press time, NUDT could not be reached for further clarification about Tianhe’s make-up, and AMD declined to offer any additional details.

If I had to speculate, I would guess that the Intel chips are Nehalem EPs and the AMD parts are FireStream 9270s. Presumably most of the FLOPS come from the GPUs. In fact, 5,000 9270s would represent 1.2 double precision petaflops all by themselves. The future Chinese CPUs are likely to be of the Godson-3 variety, which are expected to debut in 2010. Note that the Godson-3 is a MIPS architecture, but has the capability to emulate x86 instructions as well.

In addition to the sketchy details on the architecture, no mention was made of the application set the machine will be targeting. NUDT is jointly run by the Ministry of National Defense and the Ministry of Education, which gives you some idea of its areas of interest. According the university’s Web site, the institution is devoted to basic sciences, engineering, military science, management, economics, philosophy, literature, education, law, and history.

Impressive as this all sounds, Tianhe’s rather low Linpack efficiency (Rmax/Rpeak) may limit its applicability somewhat. Linpack usually represents a nominal high-water mark for what kind of performance you’re likely to get from math-intensive applications. The NUDT machine didn’t even manage to reach the 50 percent mark in efficiency — just 563 out of a possible 1206 teraflops. Most supers have a Linpack efficiency north of 75 percent, even just for vanilla GigE clusters. The new Earth Simulator in Japan boasts a 93.4 percent figure.

Undoubtedly, the problem is related to extracting Linpack FLOPS from the GPUs. Although one would think these general purpose graphics processors would excel at this type of vector math, optimal Linpack performance is also dependent on a generous cache. Modern CPUs have plenty of it, but GPUs contain only limited internal caches. That means the graphics chip would have to access the relatively slower on-board GDDR memory to refresh its data, or worse yet, go across the PCIe bus to get some more data from CPU memory. NVIDIA’s upcoming Fermi processor will be the first GPU with a true cache hierarchy (not to mention much better double precision performance), so I imagine Linpack results on this architecture should be a good deal more impressive.

In the meantime, Tianhe will represent an interesting test case for a CPU-GPU hybrid supercomputer, an architecture which is likely to become more commonplace over the next few years. It also signals China’s intention to become a bigger player in the supercomputing arena. Given the country’s huge cash reserves and the government’s willingness to invest in high-tech, there’s not much that can stop it.