AMD Begins Adding "GFX950" GPU Support To LLVM For Next CDNA Accelerator

[phoronix]

Phoronix: AMD Begins Adding "GFX950" GPU Support To LLVM For Next CDNA Accelerator

As of today the first handful of commits have landed in LLVM Git ahead of next year's LLVM 20.0 for beginning to enable the AMDGPU compiler back-end for "GFX950", the next iteration of the CDNA family for Instinct accelerators...

https://www.phoronix.com/news/AMD-GFX950-LLVM-Start

[You-]

Does that mean its still GCN?

[oleid]

Does that mean its still GCN?

They announced that they were going to merge CDNA and RDNA to UDNA, but nobody knows when this will happen exactly. Since the announcement was only in september, I doubt that it will already happen for the next hardware cycle.

[Anux]

Does that mean its still GCN?

No that is CDNA, GCN is not used on new silicon anymore.

[You-]

No that is CDNA, GCN is not used on new silicon anymore.

yes but it was never clear that CDNA was anything but GCN+++.

With this nomenclature, it seems to be confirming the case again.

GFX10xxx seems to be RDNA2, GFX11xxx seems to be RDNA3 and GFX12xxx will be RDNA4. This is still on GFX9xxx

(RDNA seemed to half (and then half again with RDNA3?) the compute performance per compute unit compared to GCN, while improving graphics throughput.)

[Anux]

yes but it was never clear that CDNA was anything but GCN+++.

As is RDNA one just got rid of all the fixed function stuff and the other cut back on compute. But CDNA is multi chiplet while RDNA and GCN aren't.

GFX10xxx seems to be RDNA2, GFX11xxx seems to be RDNA3 and GFX12xxx will be RDNA4. This is still on GFX9xxx

Those are just internal development names.

And guess what? Both will be rejoined to UDNA, hopefully bringing the chiplets to desktop GPUs to get the same effect of die reusage that Zen enjoys.

[You-]

Those are just internal development names.

Which I will suggest tell more of a story than the marketing names.

And guess what? Both will be rejoined to UDNA, hopefully bringing the chiplets to desktop GPUs to get the same effect of die reusage that Zen enjoys.

But will UDNA be based off RDNA or CDNA? Ignore the marketing, the internal engineering decisions are interesting.

(As for the marketing, I remember when the bifurcation as they called it happened, they said it then that it would eventually merge again. People seem to forget that.)

[bridgman]

yes but it was never clear that CDNA was anything but GCN+++.

More like GCN---+++ given that we removed the display engine, graphics pipeline and texture filtering but added hardware for matrix operations, more instruction overlap and new data types. Everything is debatable but we felt that continuing it GCN (Graphics Core Next) without even basic graphics functionality did not make sense.

And yes, if we had picked a different name for GCN that did not include the word "graphics" 10 years ago things would have been easier

(RDNA seemed to half (and then half again with RDNA3?) the compute performance per compute unit compared to GCN, while improving graphics throughput.)

That doesn't sound right - compute performance per CU did not really change between CDNA and RDNA. What did change is that the RDNA CUs were more complex and hence took more silicon area with no big change in compute performance.

RDNA CUs gave considerably better graphics performance but since they took more silicon area there were generally fewer CUs for any given level of graphics performance... and the "fewer CUs" led to "less compute performance" but only if you were comparing chips with comparable graphics power.