Rumor: AMD Instinct MI250X accelerator gets 110 compute units and 500W TDP

AMD may release two variants of its upcoming Instinct accelerator based on the CDNA2 architecture. One of those variants, the MI250X, would have 110 compute units and a TDP of 500W.

ExecutableFix, a leaker that more often shares correct information about AMD products, writes on Twitter that AMD is coming with Instinct MI250 and MI250X accelerators. Those chips will both be based on the Aldebaran GPU, based on the CDNA2 architecture. The accelerators also get an MCM design, or a Multi-Chip-Module. Multiple chips are combined, just like the chiplets with CPU cores in AMD’s Ryzen processors.

The Twitterer also shares the possible specifications of the MI250X variant. This accelerator would have 110 CDNA2 compute units per die. The chip would also have a boost clock of 1.7GHz and 128GB of HBM2e memory. A tdp of 500W is also mentioned and the MI250X would again be produced at 7nm. There are no specifications known for the regular Instinct MI250 yet, but presumably that model would have lower clock speeds and fewer compute units.

According to the leaker, the MI250X further achieves an F64 compute power of 47.9Tflops and FP16 and BFloat16 performance of 383Tflops. VideoCardz, in turn, writes that the chip also offers an FP32 computing power of 47.9Tflops, but ExecutableFix does not confirm this in concrete terms. To illustrate, the current Instinct MI100 accelerator achieves FP64 compute power of 11.5Tflops and FP16 performance of 185Tflops. That chip is based on the first-generation CDNA and has 120 compute units.

There have been rumors of AMD’s upcoming CDNA2 accelerators for some time now. Australia’s Pawsey Supercomputing Center announced in July that it is working on a new supercomputer with AMD MI-next GPUs. Presumably this concerns the upcoming Instinct MI250 chips. This body already reported that the accelerators would have 128GB of memory per GPU.

It is not yet known when the new accelerators will appear exactly. The AMD CDNA2 architecture is on the roadmap for 2022, but there is no concrete release date yet. The chips are intended for data centers and super computers and cannot be used as a regular video card.