AMD Launches Arcturus As The Instinct MI100, Radeon ROCm 4.0

[phoronix]

Phoronix: AMD Launches Arcturus As The Instinct MI100, Radeon ROCm 4.0

AMD is marking the SC20 virtual conference this week by launching the AMD Instinct MI100 accelerator, which is based on their CDNA architecture. Also notable and coinciding with the MI100 launch is the Radeon Open eCosystem 4.0 (ROCm 4.0) Linux release.

http://www.phoronix.com/vr.php?view=29704

[cb88]

Like literally every article gets it wrong. RDNA and CNDA are parallel evolutions away from GCN... NEITHER, = GCN. Many of the new features in RDNA are almost certainly in CDNA as well if they make sense.

[StillStuckOnSI]

I told you.
AMD is developing ROCm for non-gaming GPUs.

I don't recall any refutations of this statement, the Arcturus codename was known months ago and a new CDNA card was inevitable. The real question is whether AMD will provide an on-ramp to their data centre compute cards via ROCm support for consumer GPUs (starting from the upcoming RDNA 2). Recent discussions on this forum seem to indicate yes, but I'm witholding judgement until it shows up in the release notes

[coder]

Sounds pretty badass. TDP and pricing are the looming questions, at this point.

Meanwhile, Nvidia just launched A100 with 80 GB of HBM2E!

[coder]

Like literally every article gets it wrong. RDNA and CNDA are parallel evolutions away from GCN... NEITHER, = GCN.

Based on Michael 's coverage of the driver changes to support Arcturus, I got the distinct impression that the core ISA changed little from GCN. And that makes sense, as GCN has a reputation for being a more compute-oriented architecture (or maybe AMD even said as much, during the intro of RDNA?).

If I'm mistaken, please reference any specific technical aspects of the ISA that have changed since Vega20 (which AMD officially referred to as the last GCN product, when they announced their CDNA roadmap).

https://www.anandtech.com/show/15593/amd-unveils-cdna-gpu-architecture-a-dedicated-gpu-architecture-for-data-centers

Many of the new features in RDNA are almost certainly in CDNA as well if they make sense.

Well, not the bifucation of CU's or the halving of their SIMD width, AFAIK - RDNA's most distinguishing features!

[cb88]

Based on Michael 's coverage of the driver changes to support Arcturus, I got the distinct impression that the core ISA changed little from GCN. And that makes sense, as GCN has a reputation for being a more compute-oriented architecture (or maybe AMD even said as much, during the intro of RDNA?).

If I'm mistaken, please reference any specific technical aspects of the ISA that have changed since Vega20 (which AMD officially referred to as GCN, when they announced their CDNA roadmap).

I'd say this is *definitely wrong*... GCN ISA isn't even fixed across GCN versions. And CDNA obviously has a lot of things that GCN didn't.

[coder]

I'd say this is *definitely wrong*... GCN ISA isn't even fixed across GCN versions. And CDNA obviously has a lot of things that GCN didn't.

Not helpful. Again, I'm asking for more than PR smoke and mirrors. If you don't have any technical specifics to cite, then let's just drop it.

If precedent is any guide, AMD will eventually release a whitepaper and ISA reference, and that should settle the matter.

I never said GCN was fixed, BTW. Just that Arcturus sounded more like a normal iteration of the ISA, rather than a change on the scale of RDNA.

[zxy_thf]

Not helpful. Again, I'm asking for more than PR smoke and mirrors. If you don't have any technical specifics to cite, then let's just drop it.

If precedent is any guide, AMD will eventually release a whitepaper and ISA reference, and that should settle the matter.

I never said GCN was fixed, BTW. Just that Arcturus sounded like a normal iteration of the ISA, rather than a change on the scale of RDNA.

What a lame post.
Can't you verify it yourself by checking LLVM? https://github.com/llvm/llvm-project...U/reduction.ll

[bridgman]

The white paper is available here: https://www.amd.com/system/files/doc...whitepaper.pdf

Off the top of my head the main changes are (a) addition of matrix instructions using a new set of accumulation registers (AGPRs), (b) restructuring of cache & memory subsystem, (c) native bf16 support, (d) more and faster XGMI links, (e) 120 CU's after harvest... but it's worth reading the white paper anyways.

[cb88]

Not helpful. Again, I'm asking for more than PR smoke and mirrors. If you don't have any technical specifics to cite, then let's just drop it.

If precedent is any guide, AMD will eventually release a whitepaper and ISA reference, and that should settle the matter.

I never said GCN was fixed, BTW. Just that Arcturus sounded like a normal iteration of the ISA, rather than a change on the scale of RDNA.

Seeing as CDNA includes things that GCN doesn't even really have at all... its suffices to say that it is at worst equivalent to an updated ISA GCN 6.... really the only thing that binds GCN versions together is general architecture not the ISA. You say I am not helpful... but my point was only to point out that saying CNDA = GCN is at best a misnomer, or oversimplification. While you are the one not adding any value to the conversation merely your disgruntlement.

CNDA is *definitely* not a scaled up GCN 5 with graphics stuff removed *at all*.