ExecutableFix (opens in new tab) has shared the alleged specifications for AMD's Radeon Instinct MI250X accelerator. The HPC GPU won't compete with the best graphics cards, but it'll give Intel's Ponte Vecchio a run for its money. According to the reputable hardware leaker, AMD will launch the Radeon Instinct MI250 and MI250X, but only exposed the specifications for the latter.
The Radeon Instinct MI250X comes equipped with the Aldebaran silicon, which is built on the 7nm process node and features the CDNA2 architecture. The leaker stated that the GPU is based on the OAM (Open Accelerator Module) form factor so it's unlikely that we'll see it in a PCIe presentation.
The Radeon Instinct MI250X sports a multi-chip module (MCM) design and carries two dies that are interlinked through AMD's Infinity Fabric interconnect. It reportedly comes equipped with 110 compute units, operating at 1.7 GHz. According to the leaker, the Radeon Instinct MI250X also has 128GB of HBM2E memory at its disposal.
The Radeon Instinct MI250 will most likely come with fewer compute units and lower clock speeds. Therefore, we can assume that the non-X variant probably utilizies recycled silicon that doesn't meet the requirements for the "X" counterpart.
If the clock speeds for the Radeon Instinct MI250X are accurate, the GPU should deliver up to 47.9 TFLOPs of double-precision (FP64) performance. That's four times as much as the current Radeon Instinct MI100 (Arcturus). For comparison, Ponte Vecchio GPU has already surpassed the 45 TFLOPs mark in regards to FP32 throughput in its early silicon. The Radeon Instinct MI250X also offers up to 383 TFLOPs of half-precision (FP16/BF16) performance.
As for power consumption, the Radeon Instinct MI250X seemingly has a 500W TDP. Ponte Vecchio, on the other hand, rocks a 600W TDP, and an Intel document has revealed that OAM will require liquid cooling.