AMD has just unveiled its Instinct MI200 series of graphics cards, using MCM and offering up to 4.9 more performance than the competition.
The fastest GPU of the moment: MI200
At the AMD Accelerated Data Center event, the brand unveiled its new series of graphics cards for HPC(High Performance Computing) named: MI200
Engraved in 6nm and offering 58 billion transistors! The MI200 series uses for the first time the MCM technology with 2 die of 110 CU, that is to say a total of 220 CU (14080 shaders) and has a maximum of 128 GB of HBM2E promising a bandwidth of 3.2 TB/s, the GPU integrates up to 880 Matrix Cores, doubling the performance FP32 and FP64. All this is contained in a TDP of 500/560W.
This enticing specification comes with stratospheric performance in front of the competition, especially NVIDIA. Thus, AMD claims up to 4.9 times more TFLOPS than the A100, with 47.9 TFLOPS in FP64 and FP32 against 9.7 and 19.5 for NVIDIA. On the Matrix Cores side, we are also on significant gains with 95.7 TFLOPS in FP64 Matrix against 19.5, or 383 TFLOPS in FP16/BF16 Matrix against 312 TFLOPS in the green.
In practice, we are on the same attractive gains, with performances from 2.3 to 3.0x higher than NVIDIA’s A100 on HPC benchmarks or from 1.4 to 2.4x directly in applications.
The MI200 series will be available in two sizes and two different cards, with the MI250 featuring 208 CU and the MI250X with 220 CU in OAM for the moment and then in PCIe version.
