Comments on: The AMD “Aldebaran” GPU That Won Exascale

By: Timothy Prickett Morgan

Timothy Prickett Morgan — Sat, 11 Dec 2021 13:29:42 +0000

In reply to emerth. It does seem like an odd choice, particularly when AMD is so proud of the fact that with the Epycs, all the features are on all the time on all the SKUs--it's just what cores work and clock speeds that give the variation in the stack.

By: emerth

emerth — Sat, 11 Dec 2021 09:24:58 +0000

Hi Timothy – memory coherence disabled on all but the top SKU – is that a hard spec from AMD?

Seems a shame to disable it like that. I will probably never buy an AOM system. I might buy the MI210 PCIe card version if it had coherent memory between the two dies. If AMD thinks this is some kind of upsell it is mistaken.

By: Mark Hahn

Mark Hahn — Tue, 16 Nov 2021 22:25:54 +0000

I’d love to hear more about the coherent CPU-CPU interface. Does AMD ever still talk about HSI, or any “big picture” work along those lines? The software environment would benefit from an architecture that managed to provide uniform, coherent addressing between persistent memory, RAM, CPU, and GPU…

(Disaggregated systems are not really the same topic, since disagg people are willing to blow up latency so they can have a rack-scale “polyamorous” system)

By: Fyzhkar Raynaldy

Fyzhkar Raynaldy — Tue, 16 Nov 2021 05:15:07 +0000

I wonder why Frontier used the regular MI250 instead of the MI250X?

By: katanarama

katanarama — Mon, 15 Nov 2021 22:58:04 +0000

In reply to EC. The "software" not being mentioned is CUDA. This is going to be AMD's Xeon Phi moment.

By: Timothy Prickett Morgan

Timothy Prickett Morgan — Mon, 15 Nov 2021 15:02:11 +0000

In reply to Dave Jeffers. Definitely not, since they forked the lines into RDNA and CDNA engines, and this is a CDNA engine. They could have many similarities, obviously. But you need a different balance of stuff, and lots of ray tracing, for gaming.

By: Dave Jeffers

Dave Jeffers — Mon, 15 Nov 2021 13:27:01 +0000

Will AMD be using the same GCD in the Radeon version?

By: Current Resident

Current Resident — Thu, 11 Nov 2021 19:42:31 +0000

“Eventually, all GPUs will do their floating processing in matrix math units, and many CPUs will, too, dropping vector units when all the code is ported some generations hence.”. No; not all floating point operations benefit from the matrix units. Matrix-matrix multiplications do, and batched matrix-vector operations too, but that’s about it.

By: EC

EC — Wed, 10 Nov 2021 14:55:37 +0000

As always appreciate TPM’s perspective and deep dive on hardware architecture. Of note, the word *software* only appears once. There is an allegory in there somewhere…

By: Dan Skel

Dan Skel — Wed, 10 Nov 2021 03:54:07 +0000

Correction: “Instinct MI200 GPU accelerator from Nvidia” should read “from AMD”.