AMD launches the next-gen supercomputer-ready Instinct MI200 accelerator series
AMD has finally revealed more details on its upcoming series of accelerators, or graphics products specifically developed for world-class research, data analysis, AI and simulations. These Instinct MI200 series includes the MI250X, now touted as the world's fastest accelerator for its hyper-advanced use-cases.
The 220-compute unit (CU) MI250X has 128GB of HBM2e memory and a bandwidth of 3.2 terabytes per second (3.2TB/s, or up to 2.7 times that of its preceding generation). It is rated for up to 47.9 teraflops (TFLOPS) of FP64 performance, or as much as 95.7 TFLOPS using its new Matrix Core technology. There are 880 of these new CDNA 2-based cores, and are also rated to confer up to 383.0 TOPS of INT4/8 performance for AI-related applications.
Its new sibling the Instinct MI250 is no slouch either, with 208 CUs; an identical memory and bandwidth profile; up to 45.3 TFLOPS of FP64 performance (or up to 90.5 TFLOPS with Matrix) and up to 362.1 TOPS for INT4/8).
Accordingly, AMD claims that the MI200 line is an industry-first as it is capable of exascale (or volumes of calculations equal to 1x1018 FLOPS or more) performance. This breakthrough has won the top-end accelerators places in the next-gen Frontier supercomputer, developed in collaboration with Hewlett-Packard Electronics and the US' Oak Ridge National Laboratory and the Department of Energy.
It also integrates third-gen EPYC CPUs, which might have something to do with the AMD-made Infinity Fabric links found in the computer's new Instinct MI250X accelerators for more streamlined connections between EPYC-line silicon and other GPUs.
Frontier is now imbued with the promise of unprecedented "scientific breakthroughs" in the near future, using its leveled-up graphics power. Therefore, whatever conclusions it helps researchers of the future to arrive at, the Instinct MI200 line will be there to grab some credit.