AMD’s subsequent compute GPU could score two dies and 128 GB of memory

by

Relaxing fact: Aldebaran is the brightest well-known particular person in the Taurus constellation. It’s 44 times the size of the sun and 400 times as vivid. It’s moreover the namesake of AMD’s subsequent Intuition Accelerator.

Aldebaran, the computer phase, has regarded a few times in Linux patch notes. Its two-die manufacture modified into half-implied in February and confirmed closing month: “on Aldebaran, entirely essential die fetches legit energy data. Point out energy/energy values as 0 on secondary die,” one present reads.

It’s the essential AMD Accelerator/GPU to leverage an MCM (multi-chip module) manufacture, a formula that’s been plastered on patents for years but has entirely correct begun to materialize. Using a few chips/dies related intently permits for more scalability than in vogue monolithic designs, but moreover reduces the per-compute unit performance of a card, in particular in poorly parallelized workloads.

Nvidia is anticipated to downside Aldebaran with their very hang MCM designs per the Hopper structure subsequent year. Intel, meanwhile, is on the verge of releasing the Ponte Vecchio accelerator per the Xe HPC structure and an early MCM implementation.

Like Ponte Vecchio, Aldebaran leverages MCM to expand memory capability, patch notes imply. A present from closing week describes Aldebaran as having two dies, four UMCs per die, and eight channels related to 2 GB of HBM per UMC. Or 128 GB, in complete.

AMD’s surprisingly edgy MI100.

The patch notes moreover repeat strengthen for a “contemporary HBM2 memory form,” presumably HBM2e. If AMD is the use of SK Hynix’s contemporary HBM2e memory modules with 3.6 Gbps of bandwidth and a 4096-bit bus on every die, then Aldebaran could score 3.6 TB/s of memory bandwidth.

For comparability, Aldebaran’s predecessor, the Intuition MI100 (beforehand codenamed Arcturus, one other well-known particular person) has 32 GB of HBM2 and 1.2 TB/s of memory bandwidth. Nvidia’s A100 accelerator could moreover be configured with as a lot as 80 GB of three.2 Gbps HBM2e memory for 2 TB/s of bandwidth.

AMD CEO Dr Lisa Su has already mentioned that the CDNA2 structure, that formula Aldebaran, is slated for unlock later this year. Presumably, this could unlock as the MI200 Accelerator.

“This year, as I mentioned, we’re putting collectively our subsequent generation CDNA structure. Here’s actually a key part that enabled us to make a choice the largest supercomputer bids in the US,” Su mentioned at a conference in Could per chance presumably also fair. “… And we are in a position to be launching the following generation of that structure, actually, later this year. We’re very passionate about it. I contemplate it’s progressed extraordinarily wisely.”

Masthead by Timothy Dykes