AMD RDNA3 ISA Reference Guide Published

[phoronix]

Phoronix: AMD RDNA3 ISA Reference Guide Published

Following last month's launch of the Radeon RX 7900 series graphics cards, AMD's GPUOpen group has now published the instruction set architecture (ISA) programming guide for those interested in RDNA3 GPUs...

https://www.phoronix.com/news/AMD-RDNA3-ISA-Guide

[brucethemoose]

The matrix instructions are different than CDNA2. For instance, CDNA will take various input sizes at different precision (like 32x32x8 FP16). RDNA3 will only take a 16X16 input, but seems to support INT8 and INT4 operations unlike CDNA2. And the CDNA2 ISA paper goes into less detail about how those instructions actually get executed.


Maybe I am in over my head here, but this seems like a disadvantageous setup if AMD wants to proliferate ML usage on their cards. Nvidia's tensor cores are the same from the bottom tier GPUs all the way up to the biggest datacenter cards, which has to be a huge advantage for devs targeting both sets of hardware.

[Mahboi]

Maybe I am in over my head here, but this seems like a disadvantageous setup if AMD wants to proliferate ML usage on their cards.

I'm pretty sure they want to Xilinxify/CDNA their cards with extra chips that'll take these kinds of workloads. Maybe for RDNA4 and beyond.

[cb88]

I'm pretty sure they want to Xilinxify/CDNA their cards with extra chips that'll take these kinds of workloads. Maybe for RDNA4 and beyond.

Thats absolutely the goal as Ryzen 7040 series is already doing that with a 12TOP dedicated accelerator.... it should sip less power than doing it on the GPU also

[NeoMorpheus]

I believe they clearly stated that RDNa is for gaming and CDNa for everything else.

[shmerl]

Maybe I am in over my head here, but this seems like a disadvantageous setup if AMD wants to proliferate ML usage on their cards. Nvidia's tensor cores are the same from the bottom tier GPUs all the way up to the biggest datacenter cards, which has to be a huge advantage for devs targeting both sets of hardware.

There is some benefit in not cramming everything and your sink into a single device. Less AI hardware can mean more graphics focused hardware in the GPU.

Datacenters can indeed use non gaming cards. But even RDNA3 cards have some new AI accelerators too.

[coder]

this seems like a disadvantageous setup if AMD wants to proliferate ML usage on their cards. Nvidia's tensor cores are the same from the bottom tier GPUs all the way up to the biggest datacenter cards, which has to be a huge advantage for devs targeting both sets of hardware.

Yes, but understand that CDNA is targeted at machine learning and HPC. Therefore, its matrix cores are generalized and support higher-precision arithmetic. So, although I agree with the consequence you've identified, it's not as if they're different for no good reason. Also, CDNA and RDNA area already plenty different.

I'm guessing what AMD would probably say is that it only takes a few developers to write backend support for ML frameworks, and then it shouldn't really matter which of their GPUs you use (aside from obvious things like price, power, performance, and scalability).

[Jabberwocky]

There is some benefit in not cramming everything and your sink into a single device. Less AI hardware can mean more graphics focused hardware in the GPU.

Datacenters can indeed use non gaming cards. But even RDNA3 cards have some new AI accelerators too.

Agreed, with the exception of the MI300 where AMD crammed everything including 128GB of HBM into the CPU/APU to reduce latency. That obviously comes with a major financial cost. It also requires advanced stacking technology.

AMD has been very compute focused in their GPUs and I think it was a good choice, relative to Nvidia's tensor cores and RTX. If only the AMD ROCm drivers were better.

[brucethemoose]

Agreed, with the exception of the MI300 where AMD crammed everything including 128GB of HBM into the CPU/APU to reduce latency. That obviously comes with a major financial cost. It also requires advanced stacking technology.

AMD has been very compute focused in their GPUs and I think it was a good choice, relative to Nvidia's tensor cores and RTX. If only the AMD ROCm drivers were better.

I don't think the MI300 even supports rendering?

And on the second point: AMD is only focused on enterprise compute. They basically dropped the compute focus for individuals after Vega. Even if you are a small company who just needs a few GPUs, the major cloud vendors dont seem to offer small MI200 instances.


Mind you, ROCM has improved by leaps and bounds, but the 5000 and 6000 series are just fundamentally not great compute architectures.

[coder]

I don't think the MI300 even supports rendering?

Correct. No CDNA products have rendering hardware (e.g. texture engines, ROPs) or display outputs.

Even if you are a small company who just needs a few GPUs, the major cloud vendors dont seem to offer small MI200 instances.

Well, they make PCIe cards you can buy.

https://www.amd.com/en/products/serv...instinct-mi210

I think AMD would probably like for cloud platforms to offer instances, but maybe the providers don't see enough demand.

Mind you, ROCM has improved by leaps and bounds, but the 5000 and 6000 series are just fundamentally not great compute architectures.

I don't know what RDNA and CDNA stand for, but I'm reasonably certain the R and C are for Rendering and Compute.