AMD Instinct MI250X, AMD’s new HPC GPU specifications

ExecutableFix has published the alleged specifications of the Radeon Instinct MI250X, a variant that won’t be the top model of the upcoming series, but will compete with Intel’s Ponte Vecchio.

AMD’s Radeon Instinct MI250X with Aldebaran silicon seems to have been spotted by leaker ExecutableFix, who has shared the specifications of this graphics card for the HPC segment.

Instinct MI250X, specifications for AMD’s new HPC GPU

Radeon Instinct MI250X will use Aldebaran silicon manufactured by AMD with the 7nm process node and a CDNA2 architecture. The form factor of the graphics card will be OAM (Open Accelerator Module), which would require liquid cooling, and there don’t seem to be any plans to see this model on PCIe.

Internally, it will be an MCM GPU (multi-chip module) with two dies connected via an Infinity Fabric interconnect. The number of computing units will be 110, operating at a clock rate of 1.7 GHz. The graphics card has a capacity of 128 GB of HBM2E memory.

AMD Instinct MI250X, AMD Instinct MI250X, AMD’s new HPC GPU specifications, Optocrypto

With these specifications, the GPU should deliver up to 47.9 TFLOPs of double-precision performance (FP64). That would be four times more than the Radeon Instinct MI100 (Arcturus). It should also offer a semi-precision performance of about 383 TFLOPs (FP16/BF16).

The Instinct MI250X compute card delivers 47.9 TFLOPs of single-precision (FP32)/double-precision (FP64) compute performance and 383 TFLOPs of half-precision (FP16/BF16) compute performance, with FP64 compute performance 4.16 times that of the Instinct MI100 compute card. For comparison, the FP64, FP32, and FP16 computing power of the Nvidia A100 80GB version are 19.5 TFLOPs, 156 TFLOPs, and 312 TFLOPs respectively, so I expect the two sides to meet again soon.

As for power consumption, this model should be rated at 500W TDP, 100W below Intel’s new Ponte Vecchio.

AMD intends to launch the Radeon Instinct MI250X and MI250 models, the latter as a smaller sibling of the X version with fewer processing units. Unfortunately, the leak only refers to the X model.

When will we see them in action? The release date is completely unknown at the moment, but it could be in the first half of next year.