When AMD last published roadmaps, they included the CDNA 2 computing architecture as a product of 2022. Nevertheless, RedGamingTech’s editors claimed this winter that products built on AMD’s second-generation computing architecture could appear as early as 2021.
She probably seemed right. At least the work procedures suggest this: Samples of the MI200 accelerator (which according to most sources is based on CDNA 2) already exist in such an advanced state that AMD does not try to put them into operation in laboratories, but development is already at the stage when accelerators are tested by partners. companies. Specifically, it runs in the HPE Cray EX class supercomputer accompanied by a semi-custom server processor Trento built on architecture Zen 3. According to leaker Bondrew, it is specifically a Shasta-type system and is equipped with an MCM version of the accelerator, which carries two pieces of silicon in one housing. There should also be a single-silicon version.
Finished product, ie systems with Trento and CDNA 2 / MI200, are to be (commercially available from a roadmap published by leaker Komachi) commercially available this year.
According to ISA, CDNA 2 / MI200 should support full-rate double-precision, ie calculations with FP64 accuracy in native performance. This will probably be the first GPGPU solution to offer this convenience (this means, among other things, that compared to the previous generation, the FP64’s performance will be 2x higher with the same number of stream processors). The architecture according to the designation of the GFX90A instruction set corresponds to the development branch derived from the GCN optimized for computational deployment.
While AMD will probably actually distribute this year according to the roadmap and the current state of preparations, in the case of the competing Xe-HPC it does not look completely optimistic yet. It was supposed to appear for the first time at the end of this year in the completed Aurora supercomputer, but some sources since the end of last year suggest that it will not go so fast. Aurora is to be built on Intel server processors Sapphire RapidsAccording to several sources, they were postponed to 2022. As far as the Xe-HPC accelerators are concerned, it would be quite unusual for only three quarters to escape between the creation of the first sample and the supply of mass-produced pieces, including assembly and start-up of the supercomputer.
Although it was originally reported that Aurora with Intel Xe-HPC and Sapphire Rapids will be ready in 2021 and AMD MI200 accelerators with CDNA 2 architecture will not be delivered until 2022, all indications now are that AMD and Intel will swap these data.