After being banned from exports by the United States, the Guangzhou Supercomputing Center upgraded the Tianhe 2A supercalculation using a domestic Matrix-2000 acceleration card, and improved its performance from 54.9PFLOPS to 97.9PFLOPS.
Tianhe 2 is a supercalculation at the Guangzhou Supercomputing Center in China. It uses the Intel Xeon E5-2692 12 core processor and the Xeon Phi 31S1P acceleration card, with a total of 3 million 120 thousand cores and a total power of 17808 kW. The theoretical performance is 54.9PFLOPS (tens of millions of billion), Linpack peak performance 33.86PTFLOPS, from June 2013 By June 2016, its performance has not changed, and has won six TOP500 Championships.
Tianhe 2 had a late escalation, but in 2015 the U.S. government banned Intel and other companies exporting high performance computing chips to China's four supercomputing centers, so that the Intel Xeon Phi acceleration card, Tianhe 2, would not be able to use high performance chips from American companies.
In September 2017, the Guangzhou Supercomputing Center announced the upgrading of the Tianhe 2 supercomputing system by the end of the year, replacing the original Intel Xeon Phi accelerator with the domestic accelerator Matrix 2000.
The upgraded Tianhe No. 2 is called Tianhe 2A, and the name of Tianhe 2A is also used in the previous reports at home and abroad, but its real upgrade is the end of last year. This is the true Tianhe 2A,Floating point performance has been upgraded from 54.9PFLOPS to 94.97PFLOPS.
Judging from the upgrades, the Tianhe 2A is not just as simple as using domestic accelerators instead of Intel accelerators.The network structure has also been upgraded from the original 10Gbps to 14Gbps, the delay from 1.57us to 1US, the memory capacity from 1.4PB to 3.4PB, the storage capacity from 12.4PB to 19PB, the bandwidth doubling to 1TB/s, and the power consumption from 17.8MW to 16.9MW, the energy efficiency is greatly improved.
The key to the Tianhe 2A upgrade is the above Matrix 2000 accelerator, which uses domestic chips, and the architecture and source are as mysterious as ever. After all, it is related to the National Defense Department.
You can find this information on the Internet.This chip is produced by Tianjin Mai Chuang, each acceleration card uses 4 Matrix 2000 chips, each Matrix 2000 consists of 128 cores, frequency 1.2GHz, each cycle can perform 16 dual precision operations, the processor's peak performance is 2.45TFLOPS.
The Matrix 2000 processor's core architecture is reminiscent of the Deuteronomy processor used by the light of Taihu, but the latter is based on the Alpha architecture, and the Matrix 2000 architecture is not ARM.
Considering the processor roadmap published before the National Defense Department, the integer architecture should be ARM, but the vector unit is still the domestic magic change. The official data mentioned in the official data is the custom 256bit VFU vector unit.
In addition,The TDP power consumption of Matrix 2000 processor is 240W, the encapsulation area is 66x66mm, the manufacturing process is unknown, but from the time of release, it is probably produced by 28nm node.