Succeeding the MI100, the MI200 which should appear at the next AMD earning call, would be a power monster in FP16 and FP32.
47.9 TFLOPS of performance in FP32
The Radeon Instinct MI200 is the next professional GPU from AMD, succeeding the MI100 based on the CDNA architecture. The MI200 will be in CDNA2 (still derived from Vega). This GPU, which is intended for computing and will equip the next supercomputer Exascale Frontier, will compete with the Nvidia A100 and the Intel Ponte Vecchio. Today, the leaker @ExecuFix gives us new information on the characteristics and performance of this card.
The card engraved in 7nm, would use a MCM technology with 2 die of 110 CU each, giving a total of 220 CU (14080 shaders) all clocked at 1.7 GHz. The bug would allow monstrous performances with 47.9 TFLOPS in FPP32 and 47.9 TFLOPS in FP64. It would even go up to 95.8 TFLOPS with the Matrix Core, all accompanied by 128 GB of HBM2E memory.
In comparison, the A100 of Nvidia delivers 19.5 TFLOPS in FP32 and 9.75 in FP64. On the Intel side, the Ponte Vecchio GPU reaches 45 TFLOPS in FP64 and 32. The MI200 would crush the competition in terms of raw performance.
Of course, this needs to be confirmed by AMD, which could speak about Frontier and the components that make it up (including the MI200), October 26 during their earning call of the third quarter.
