China reveals home-grown supercomputer chips after Intel x86 ban
The speeds and feeds of the Middle Kingdom's custom HPC DSPs
ISC 2015 Computer design guru Yutong Lu has outlined her blueprints for China's homegrown supercomputer chips following a US ban on Intel processor exports to the Middle Kingdom.
In April it emerged that Uncle Sam had forbidden Intel from shipping high-end Xeon and Xeon Phi parts to China's defense labs and other areas of its supercomputing industry. These components are more or less essential in building powerful number-crunching machines, and China had to come up with a workaround. Sooner than expected, the nation's boffins have done just that.
Dr Lu, who leads the design of China's Tianhe supercomputers, said homegrown digital-signal processors (DSPs) will power the upgrade to the Tianhe-2A super, our sister website The Platform reports. Dr Lu revealed the development at the International Supercomputing Conference in Germany on Wednesday.
The boosted Tianhe-2A is due to go live before the end of 2016, and is apparently expected to perform 100PFLOPs – 100,000 trillion calculations per second – at its peak. It will, according to Dr Lu, consume up to 18MW of power, pack about three petabytes of system RAM, and use Intel Xeon E5-2692 processors from the Tianhe-2 plus the new homegrown accelerators.
Today's Tianhe-2 – the world's most powerful publicly known supercomputer – uses a mix of E5-2692 CPUs and Xeon Phi accelerators. Essentially, the 2A will use the China-crafted DSPs instead of the Phis, alongside the Xeon E5 processors, it appears. The Tianhe-2A will be built from 18,000 nodes, and run off a 30PB file system, we're told.
The Matrix2000 DSPs are 1GHz 64-bit chips, they draw 200W of power, and can perform 4.8TFLOPS of single and 2.4TFLOPS of double precision math. They are interfaced using x16 PCIe 3.0 links, and performance-wise, give GPUs and the Phi a run for their money.
"The Tiahne-2 machine (and its eventual successor sporting the DSP accelerators) is housed at the National University of Defense Technology (NUDT) in China," The Platform co-editor Nicole Hemsoth reports.
"As a side note, the fact that the center is a major defense and Chinese national security research center and DSPs are being leveraged in high-performance systems as co-processors with all of the necessary software stack and programming environments to support it means this is likely something that has been in the works for military and defense systems within China.
"DSPs are frequently used in embedded military applications, including remote sensing, radar, and other activities — and Lu did tell the group this week that NUDT has had extensive experience with DSPs."
Lobbing these specialized chips into high-performance computers is not a new concept: you can program them to perform floating-point calculations in hardware, which means the math is done extremely quickly, and they can execute code over multiple cores [Unleashing DSPs for General-Purpose HPC PDF]. Normally, you'd find DSPs in systems doing audio and signal processing, video compression, speech recognition, and so on.
For a full rundown of the Tianhe-2A's DSP architecture, check out The Platform's coverage. ®
In other conference news
- SGI is touting new UltraViolet shared-memory systems tuned to run SAP's HANA database and big iron workloads.
- Intel and HP have entered an alliance to get serious about supercomputer hardware. Intel is also still talking about its Knights Landing Xeon Phi coprocessor for acceleration heavy workloads, and Omni-path interconnect.
- And the list of top-500 most powerful publicly known supercomputers has been revised: China is still top of the charts, but the US has the most machines on the list.