Shulou(Shulou.com)12/24 Report--

Thanks to CTOnews.com netizens Baa Yang, soft media users 1520111 clue delivery! CTOnews.com November 26, SC (International Conference for High Performance Computing, Networking, Storage, and Analysis) is a top-level meeting in the field of high-performance computing and architecture.

At the SC23 just held, China showed the world the brand-new Shenwei SW26010-Pro processor, whose performance is four times higher than that of the previous generation. This is a series of processors with completely independent intellectual property rights in China, which adopts independent instruction set (extended based on Alpha).

According to reports, Shenwei Taihu Light, which is based on the Shenwei SW26010 processor, currently ranks 11th in the Top 500 supercomputer list, while the supercomputer using the new processor ranks second, second only to the Frontier of Oak Ridge National Laboratory in the United States.

▲ SC23 PPT, image source: Rongfen LinSW26010-Pro consists of 6 core groups and 1 protocol processing unit (PPU). Each core group contains 64 computing processing elements, with a total of 384 cores, compared with only 4 core groups in SW26010.

In addition, the memory controller it supports has been upgraded from DDR3 to DDR4-3200; and each core group has its own memory controller with 16 GB memory and a theoretical bandwidth of 307.2 GB/s. A chip can access 96 GB of main memory, a significant improvement over SW26010's 32 GB implementation.

According to Chips and Cheese, SW26010 Pro C runs faster than its predecessors (CPE 2.25GHz,MPE 2.10GHz, compared with 1.45GHz), coupled with improvements in the new 64-bit RISC microarchitecture and system architecture, its FP64 computing throughput has more than quadrupled.

The data show that the maximum FP64 throughput of each SW26010 Pro is 13.8 TFLOPS. From the perspective of theoretical calculation of throughput, the new supercomputer is very powerful. Although SW26010-Pro is similar to Fugaku A64FX at the node level, our supercomputers can provide higher FP64 throughput while using fewer chips. The comparison of CTOnews.com summary data is as follows:

CPU computing core FP64FP32SW26010-Pro38413.8 TFLOPS27.6 TFLOPSSW260102562.9 TFLOPS5.8 TFLOPSA64FX483 TFLOPS6 TFLOPSMI250X (single GCD) 11023.9 TFLOPS23.9 TFLOPS | 47.8 TFLOPS (packed) however, its memory bandwidth does not match the computing throughput, so memory bandwidth becomes its performance bottleneck.

SW26010-Pro tries to solve the problem by increasing the cache capacity to 256KB (64 KB for SW26010), but in the absence of a proper L2 cache, only 256KB cache per CPE is obviously not enough, so there are still major performance bottlenecks on both processors. At the same time, a dual-channel DDR4-3200 (51.2 GB/s) memory subsystem is not enough for 64-core processors with 512-bit vector FPU and up to 16 FP64 FLOPS per cycle.

Supercomputing system node HPL-MxP performance (Eflop/s) theoretical FP64 throughput (Pflop/s)

New generation Shenwei 107136 Shenwei SW26010-Pro5.0481481.05 Shenwei Taihu Lake Light 40960 Shenwei SW26010? 121.6 Fuyue 158976 A64FX2537.21Frontiers9408 (64c AMD Xiaolong 7A53 + 4x AMD MI250X) 9.95071821.84

(only GPU is 1679.62) related reading:

"the latest TOP500 supercomputing rankings have been announced: Frontier remains number one in the world, Aurora and Eagle are airborne."

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.