Il leaker ExecutableFix has spread new rumors about the future AMD accelerators of the Instinct series. For some time there have been rumors of the arrival of the new offer within the year, but now it seems that we will immediately see two models calledMI250 e MI250X based on a 7-nanometer GPU calledAldebaran based on the new architectureCDNA 2.
The MI250X should have a total of 110 Compute Unit (7040 stream processor) with a frequency of 1,7 GHz so i 128GB of HBM2E memory, as per previous rumors. The throttle should have a TDP of well500Wand guarantee a power of approx 48 TFLOPSwith both double precision (FP64) and single precision (FP32) calculations, while it should touch well383 TFLOPSwith FP16 (half precision) calculations.
With regard to MI250, the GPU should have fewer active Compute Units (and perhaps lower frequencies), but at the moment the exact specifications have not yet been leaked. Both should therefore be based on the Aldebaran GPU, the first from AMD (indeed the first ever) based on an MCM (Multi-Chip Module) project, where two GPUs are placed on the same package and work together (a sort of CrossFire at much more specific, said in a cheesy way).
AMD’s two accelerators should face the supercomputer and artificial intelligence market with Intel Ponte Vecchio (Xe HPC) and NVIDIA H100 (Hopper), solutions that should only debut next year. Meanwhile, AMD has already started to design MI300 accelerators always on MCM design and has announced plans to improve the energy efficiency of EPYC CPUs and Instinct accelerators by 30 times by 2025.