Visit additional Tabor Communication Publications
November 01, 2011
If anyone wasn't taking China seriously as a contender for supercomputing supremacy, such doubts should have been dispelled last week when the New York Times reported that the nation has deployed its first petascale supercomputer built with domestically produced CPUs. And it's not just the processors that were homegrown. Based on a presentation delivered last month at the China's Annual Meeting of National High Performance Computing, most of major components of the new machine were designed and built with native engineering, including the liquid cooling technology, the system network, and the software stack.
As we recapped last week, the Sunway BlueLight MPP, installed in September at the National Supercomputer Center in Jinan, is being powered by 8,704 ShenWei SW1600 processors. The resulting machine delivers just a over a petaflop of performance, with a Linpack rating of 796 teraflops. That will probably place it somewhere between 15th and 20th place on the upcoming TOP500 list, assuming the engineers at Jinan sent their submission in on time.
Impressively, its power consumption of just one megawatt will make it one of the more power-efficient of CPU-only supercomputers in the work. Running Linpack, BlueLight delivers 741 megaflops/watt, which would place it in the top ten of the current Green500, a list that ranks the energy efficiency of supercomputers.
Perhaps even more impressively, this was all accomplished with CPUs built on 65nm process technology, which is two generations behind what can be had at most of the major fabs today. According to the presentation last week, the domestic ShenWei chip is a 16-core, 64-bit RISC processor running between 0.975 - 1.2 GHz. Assuming a frequency of 1.1 GHz, the CPU will can deliver a peak double precision floating point performance of 140.8 gigaflops. Note that if the 8,704 CPUs were running at that speed, the machine would actually deliver 1.2 peak petaflops, not the claimed 1.07 petaflops. Apparently the supercomputer is equipped with processors clocked at the lower end of their frequency range.
Digging a little deeper into the specs, the CPU is a four-issue superscalar design with two integer and two floating-point execution units. The integer unit has a 7-stage pipeline, while the floating point unit is implemented as a 10-stage pipeline. The system bus is 128 bits wide.
As is the case with most CPUs nowadays, the chip contains an integrated DDR3 memory controller. It feeds the 16 cores at a rate of up to 68 GB/second, using four memory channels. Each of the machine's CPUs is directly connected to 16 GB of memory, although the ShenWei's maximum memory reach is a whopping 1 TB (and 8 TB for virtual memory).
The chip also contains Level 1 and Level 2 caches --- 8 KB each of instruction and data for L1, and 96 KB for L2. Those are rather small by modern CPU standards, but considering the relatively large geometries of 65nm transistors, there probably wasn't room for both large caches and lots of cores. In this case, the chip architects opted to maximize core count.
Design of the ShenWei microprocessors is being attributed to the Jiangnán Computing Research Lab, with support from the Shandong government. The chips themselves are being fabbed by "a company in Shanghai," which plans to moves from the current 65nm process node to 45nm. According to the Wikipedia entry on the ShenWei, this is the third generation of the architecture.
The CPUs are rather densely packed in the BlueLight system. Each 1U box crams together four dual-socket motherboards, which is about two to four times the density of a typical design. Normally that would make for an uncomfortably hot enclosure, so to compensate, the system is entirely water cooled. From the pictures in the presentation, it looks like piped liquid is run through the motherboard to maximize heat dissipation.
Each node -- what they refer to as a super node -- consist of 256 CPUs (4,096 cores) and 4 TB of memory, providing 32.7 teraflops of peak performance. Intra-node communication is supported by a high-speed backplane, which delivers 1 terabyte/second of bandwidth.
The system network is the most conventional part of the machine, being based on QDR InfiniBand. In this case, the engineers built custom-made 256- and 324-port switches, and outfitted the connections with optical fiber. The network is a fat-tree topology and is designed for optimized routing as well as dynamic fault tolerance. It's not clear if Mellanox or QLogic components are in the mix here, but no mention was made of third-party switch ASICs or NICs.
The software stack is attributed to Sunway, which has provided the "virtualization" management, a parallel operating system, the parallel file system, the compiler for the ShenWei CPUs, multicore math libraries, and a Java support platform. Compiler support includes the usual suspects: C, C++, and Fortran, as well as UPC and OpenMP. The requisite MPI library rounds out the software stack.
With the ShenWei CPU, China has begun the process of edging out foreign-built processors with its own designs. The BlueLight machine first supercomputer on China's TOP100 list with homegrown CPUs. At it stands now, 85 of those systems use Intel processors, with the remaining 14 using AMD parts. It's clearly China's intent to reduce, or perhaps even eliminate entirely, its dependence on processors designed outside its borders -- at least for its HPC needs.
In aggregate, the Chinese have built a what appears to be world-class supercomputer, designed and built without the help of any US-based chipmakers or system vendors. The Japanese, of course, accomplish this a fairly regular basis, the latest example being the K supercomputer at RIKEN. By contrast, Europe possesses only an incomplete domestic HPC industry, with system vendors like Bull relying on exogenous CPUs, interconnects, and other components. For China, a relative newcomer to the world of high-end HPC, designing and building a domestic supercomputer is a major achievement.
Should vendors be worried? Certainly chipmakers like Intel, AMD, and NVIDIA should view this development with some trepidation. Likewise for HPC system vendors such as IBM, HP, Dell and others. China is a large and growing market for high performance computing infrastructure, and if they decide to take a homegrown approach to HPC technology, that could translate into hundreds of millions of dollars per year in lost revenue for these US-based companies.
As far as the broader picture of US (and European) competitiveness in HPC capability, there is also reason for concern. A number of industry insiders believe the Chinese are determined to beat the US and other nations in the race to exaflops. Convey co-founder and chief scientist Steve Wallach is one such individual. According to him, the dense packaging, impressive performance per watt metrics, and water cooled technology of the BlueLight system are signs of serious engineering prowess on the part of the Chinese engineers.
"This is ground-up design," Wallach told HPCwire. "They own the technology, and that's the key."
More importantly, he believes the technology can scale more easily than mainstream products being offered in HPC today. In particular, if the Chinese catch up (or outsource) to more advanced fab technology, the ShenWei processors could be quite formidable. According to him, compared to a 65nm die, 32nm technology would provide four times the available silicon real estate, freeing the ShenWei designers to add more cache -- something Wallach believes is a weakness in the current design.
A more obvious advantage is that, rather than relying on commodity processors and commercial clusters, the Chinese government seems willing to develop processors and systems targeted specifically to HPC. The Japanese government has done this to some extent with the aforementioned K machine and the NEC vector machines, but in the US and Europe, there is no direct government support to fund HPC processors, and only piecemeal support from various agencies to design and build advanced supercomputing systems.
In that sense, the Chinese can exploit their considerable financial resources to outrun the competition if they choose to do so. And if the new ShenWei processor and the BlueLight system is an indication of a systematic strategy, then the Chinese have already made that choice.
May 16, 2013 |
When it comes to cloud, long distances mean unacceptably high latencies. Researchers from the University of Bonn in Germany examined those latency issues of doing CFD modeling in the cloud by utilizing a common CFD and its utilization in HPC instance types including both CPU and GPU cores of Amazon EC2.
May 15, 2013 |
Supercomputers at the Department of Energy’s National Energy Research Scientific Computing Center (NERSC) have worked on important computational problems such as collapse of the atomic state, the optimization of chemical catalysts, and now modeling popping bubbles.
May 10, 2013 |
Program provides cash awards up to $10,000 for the best open-source end-user applications deployed on 100G network.
May 09, 2013 |
The Japanese government has revealed its plans to best its previous K Computer efforts with what they hope will be the first exascale system...
May 08, 2013 |
For engineers looking to leverage high-performance computing, the accessibility of a cloud-based approach is a powerful draw, but there are costs that may not be readily apparent.
05/10/2013 | Cleversafe, Cray, DDN, NetApp, & Panasas | From Wall Street to Hollywood, drug discovery to homeland security, companies and organizations of all sizes and stripes are coming face to face with the challenges – and opportunities – afforded by Big Data. Before anyone can utilize these extraordinary data repositories, however, they must first harness and manage their data stores, and do so utilizing technologies that underscore affordability, security, and scalability.
04/15/2013 | Bull | “50% of HPC users say their largest jobs scale to 120 cores or less.” How about yours? Are your codes ready to take advantage of today’s and tomorrow’s ultra-parallel HPC systems? Download this White Paper by Analysts Intersect360 Research to see what Bull and Intel’s Center for Excellence in Parallel Programming can do for your codes.
In this demonstration of SGI DMF ZeroWatt disk solution, Dr. Eng Lim Goh, SGI CTO, discusses a function of SGI DMF software to reduce costs and power consumption in an exascale (Big Data) storage datacenter.
The Cray CS300-AC cluster supercomputer offers energy efficient, air-cooled design based on modular, industry-standard platforms featuring the latest processor and network technologies and a wide range of datacenter cooling requirements.