AMD: 128GB of HBM2E memory for Aldebaran dual-GPU accelerator?

For some time there have been increasingly insistent rumors of the arrival of a new one Instinct throttle of AMD for the sector HPC, characterized by two GPUs on one package. This dual-die solution, therefore belonging to the group of projects MCM (multi-chip module) that we will see becoming more popular in the future, known by the code name Aldebaran and according to increasingly concrete rumors, he should have well on board 128GB of HBM2E memory, an unprecedented amount if we think that Nvidia A100 stops at 80 GB.

A clue to this specification comes from Setonix, an HPE Cray 50 petaflops supercomputer that will serve to manage data from the Square Kilometer Array (SKA), the largest radio telescope project in the world. The specifications of this supercomputer indicate the presence of more than 200 thousand AMD Milan cores (EPYC 7003 Zen 3) and above all of over 750 “AMD MI-Next” accelerators with 128GB per GPU. Since there is no Instinct solution with this memory capacity today, it is very likely Aldebaran.

Instinct MI200 should therefore be hidden behind MI-Next. A project that should represent a clear step forward compared to the current MI100, a solution based on CDNA architecture with 120 Compute Units (7680 stream processors) and only 32 GB of HBM2 for a bandwidth of 1.2 TB / s.

GPU diagram of the new idealized AMD accelerator by an enthusiast and leaker

In the case of the rumored MI200 we are talking about 128 CU per die, then 256 CU, although not all units may be active – some talk of “only” 224 active CUs. In any case, it would be a very important leap forward. The fact that it is based on CDNA architecture 2, so in addition to the greater number of cores and memory, the architecture itself should also guarantee an important performance boost.