Chinese CPUs are just the beginning… Middle Kingdom to go for GPUs too
China is enjoying increasing success with its own CPU developments, even though for now it is confined to the internal markets mostly. However, they are already undertaking the next step, and preparing to move into the GPU arena.
Earlier this year, we covered the raising phenomenon of Chinese CPUs in quite a detail [part1, part2, part3], including the main high end lines of Loongson (MIPS) and Shenwei (Alpha), as well as over a dozen of ARM licensees there. Then there was Icube with their UPU, a truly fused CPU plus GPU at the core and register level.
There's more to it: at the high end front, besides the academic-government MIPS and purely military, for now, Alpha, our friends in the north also developed their own Fengtian SPARC compatible processors. For now, these are not used as the main CPUs, but as I/O processors in some large supercomputers to help manage the ultrafast interconnects, twice the Infiniband QDR speed, that the country uses in their largest machines such as Tianhe. These interconnects have the SPARC CPU to manage the general I/O, and a NPU, network interconnect processor, with most of the protocol overhead hardwired. Therefore, the main Xeon or other CPUs in the cluster don't need to be bothered handling the interconnect I/O overhead, raising the total real system performance.
However, don't be surprised to see these SPARC processors go further in the near future, to become main CPUs in specific uses, even with much faster FP units, for instance – the current units are well multithreaded and multicore, as far as I understand, but without SIMD FP yet, which should appear in the next generation.
Even more interesting is the initial dedicated GPU design effort in China, in fact at least two of them: one is expected to come from the same team in the Tianhe supercomputer, and the other one is coming from far north, in Harbin. While very little is known about either of them for now, there is one common point I understand is valid for both: they will focus on OpenGL and Compute GPU use, not consumer-grade DirectX gaming – China is seemingly not interested in getting even more of its flock addicted to games, but would gladly use the SIMD or vector-like GPU math capabilities in its computing efforts, though. That might be quite commendable, after all.