China's National Supercomputer Center (NUDT) introduced their Tianhe-3 system as a prototype back in early 2019—at the time it had been tested by thirty local organizations. Notable assessors included the Chinese Academy of Sciences and the China Aerodynamics Research and Development Center. The (previous generation) Tianhe-2 system currently sits in a number seven position of world-ranked Supercomputers—offering a measured performance of 33.86 petaFLOPS/s. The internal makeup of its fully formed successor has remained a mystery...until now. The Next Platform believes that the "Xingyi" monikered third generation supercomputer houses the Guangzhou-based lab's MT-3000 processor design. Author, Timothy Prickett Morgan, boasted about acquiring exclusive inside knowledge ahead of international intelligence agencies—many will be keeping an eye on the NUDT, since it is administered by the National University of Defence Technology (itself owned by the Chinese government).

The Next Platform has a track record of outing intimate details relating to Chinese-developed scientific breakthroughs—the semi-related "Oceanlight" system installed at their National Supercomputer Center (Wuxi) was "figured out" two years ago. Tianhe-3 and Oceanlight face significant competition in the form of "El Capitan"—this is the USA's prime: "supercomputer being built right now at Lawrence Livermore National Laboratory by Hewlett Packard Enterprise in conjunction with compute engine supplier AMD. We need to know because we want to understand the very different—and yet, in some ways similar—architectural path that China seems to have taken with the Xingyi architecture to break through the exascale barrier."Prickett Morgan reckons that the Chinese duo are performant enough to sit atop global supercomputer rankings: "rumored performance of Xingyi (AKA Tianhe-3) at NSC Guangzhou - 2.05 exaflops peak and 1.57 exaflops sustained on High Performance LINPACK - that it was the most powerful machine yet assembled on Earth. And we also reminded everyone that the Oceanlight system at NSC Wuxi was the second most powerful machine on the planet, with 1.5 exaflops peak and around 1.22 exaflops sustained on LINPACK." The report included a Top 30 supercomputer rankings chart (see above). The upcoming US-based solution could pip both to the post: "We think there is a very good chance that El Capitan will come in at 2.3 exaflops peak when it is fully fired up, hopefully by the June 2024 rankings. You can bet Elon Musk's last dollar that the premiere supercomputing lab in the United States will be getting all the help it needs to beat what we think the two Chinese machines can do."

Past reports had the Tianhe-3 (2019) prototype supercomputer utilizing an ARM-based "Phytium" CPU (FT-2000+), and "(an) updated Matrix-2000+ DSP accelerator, which is also known as the MT-2000+." The full fat Tianhe-3 system apparently leverages a set of very potent internal hardware, with some significant design revisions—according to The Next Platform: "MT-3000 is not an accelerator, but a hybrid device with CPU and accelerator compute as well as three different kinds of memory, two of which are located in the compute complex. So it is akin to the AMD "Antares" MI300A CPU-GPU hybrid that is going into El Capitan than it is like the discrete CPU-GPU systems we see pushing the flops in AI and HPC systems these days. The MT-3000 is its own animal, and you might assume that it uses a chiplet packaging architecture given that Semiconductor Manufacturing International Corp (SMIC), the indigenous Chinese foundry, probably could not cram enough transistors into a 14 nanometer process to make a monolithic die. But, then again, maybe this is a 10 nanometer or even a 7 nanometer device. If NUDT doesn't care about cost, then the yield can be terrible so long as SMIC can find tens of thousands of good MT-3000 parts to make the system."

A lot this information was extracted from paywalled articles—mostly from the "MT-3000: A Heterogeneous Multi-Zone Processor For HPC" abstract.

MT-3000"With high performance computing (HPC) continues evolving, high performance microprocessor, which is the key building block of super computer, becomes the jewel in the crown of HPC. To this end, we propose MT-3000, a heterogeneous multi-zone processor for HPC, which is entirely designed and implemented by National University of Defense Technology. MT-3000 contains 16 general purpose CPUs, 96 control cores and 1,536 accelerator cores, which are grouped into a general purpose zone and an acceleration zone. The acceleration zone is further divided into four clusters. Through sophisticated designs of such multi-zone organization, interconnection, and memory subsystem, MT-3000 achieves 11.6 teraflops double precision performance and 45.4 gigaflops/watt power efficiency when operating at 1.2 GHz. Based on the MT-3000 chip, a supercomputer prototype with nearly 12 petaflops peak performance is implemented, achieving 80 percent computation efficiency for LINPACK. The possibility of a larger scale supercomputer construction based on MT-3000 chip is also elaborated in this paper."