RDNA4

The optimist in me says combined R&D, software efforts, userbase etc should make this a good thing long term, the cynic in me says AMD sees their gaming efforts fall off a cliff and want to consolidate Radeon and give them HPC/AI scraps.. Hopefully it's a good thing long term. Wave32 or 64?



UDNA 6 then? I'm reasonably sure there have been rumours about AMD doing this for RDNA 5 too but been a while since I've bothered checking any

They want a unified programming model, like Nvidia with CUDA and Intel w/OneAPI. And sure, maybe RDNA5 gets matrix multiplication units from CDNA4, or CDNA4 gets a compute focused version instead of AI that borrows RDNA's compact dual floating point units, cause why double up the design of the same IP?

But this isn't some major business reorg. Zen cores have already shown up on CDNA with MI300a, AMD has already borrowed from one IP for another (memory bus dies), this is just continuing that trend.
 
Last edited:
IMHO, there have been too many subvariants of architecture floating around lately to really focus support on. Hope that the the line drawn between CDNA (AI-/HPC-Datacenter!) and RDNA (everything consumer/pro) is the right one to erase. There's quite a bit of stuff in the chips, that is of no value to the respective other. So hopefully, it is modular enough so that putting together the IP blocks is no major source of trouble, time and money for them.

From a Software-Company-to-be's perspective definitely the right move. Think about FX 5200.
 
Perhaps the CDNA-RDNA split was a pragmatic compromise back then — between the desire to experiment with GPU architectures, and the need of a stable derisked target for Instinct wins/growth.

Now having gone through some generations on both tracks, the classic GCN lineage of CDNA is probably running into a dead-end for innovation. GCN CU’s clever trick — of round robin issuing to each of the four fixed 4-cycle VALU pipelines every 4 cycles — is also the tightest shackle on itself.

The revolutionary directions of CDNA are most probably *pikachu shock* the RDNA architectural changes, if not the implementation (*cough* VOPD) then at least in spirit. e.g., single cycle issue, Wave32, hardware dependency resolution, etc.
 
Last edited:
Guessing everyone's seen the PS5 Pro reveal, one of the major improvements mentioned is "Advanced Ray Tracing: We’ve added even more powerful ray tracing that provides more dynamic reflection and refraction of light. This allows the rays to be cast at double, and at times triple, the speeds of the current PS5 console."


This would pretty much confirm its RDNA4 IP and bodes well for RDNA4 competitiveness in ray tracing.
 
This would pretty much confirm its RDNA4 IP and bodes well for RDNA4 competitiveness in ray tracing.
With limited production run the PS5 Pro might be the ideal testing ground for the new ray tracing IP assuming a completely revamped architecture.
 
With limited production run the PS5 Pro might be the ideal testing ground for the new ray tracing IP assuming a completely revamped architecture.

What would testing the PS5 Pro accomplish? IP Testing would have been done in simulation/silicon test vehicles a long time back and the IP is already in the RDNA4 silicon which taped out months ago.
 
With limited production run the PS5 Pro might be the ideal testing ground for the new ray tracing IP assuming a completely revamped architecture.
Huh? Testing on something then immediately release the real thing in 2 months?

Do you think RDNA4 is developed in weeks and TSMC can tape-in tape-out in days or something? Navi48 and Navi44 are both in production by now.
 
The PS5 Pro will probably sell more than the entire RDNA4 discrete graphics card lineup combined. It obviously won't be used whether RDNA4 actually works - AMD did that long ago - but it will encourage developers to optimize RT code for the new AMD hardware and hopefully provide feedback to AMD (or to Sony, who can pass it along to AMD).
 
Back
Top