RDNA5 Speculation

[Frenetic Pony]

RDNA4 is out in less than 6 months! And we've got info on RDNA5 already:

DGF: A Dense, Hardware-Friendly Geometry Format for Lossily Compressing Meshlets with Arbitrary Topologies (hardware compressed mesh format that's raytracing friendly, already praised by lead designer of Nanite)

"the largest technical leap you will have ever seen in a hardware generation" (PR, but multiple leaks point towards Xbox going w/AMD again)

"RDNA5 to feature entirely chiplet arch" (according to patents)

"Medusa Point Zen 6/RDNA5 in 2026" (Zen 6/RDNA5 already scheduled for laptops/mobile in 2026)

Supposedly targeting late 2025 for desktop versions, though who knows if that date will be hit or if CES - Computex 2026 is more likey. Still, that's a heck of a lot info, implied and leaked (yes Medusa as a codename has come up more than once and looks to be guaranteed as the actual codename). Feels like enough to start a thread.

Other leak points towards some sort of increased AI instruction throughput. Maybe it's that block FP16 (an entire set of matrix multiplication instructions share the same exponent) that AMD is so fond of (already shipping in XDNA2, confirmed support in MI350)

 

[Albuquerque]

But can it fold proteins?

 

[Erinyes]

RDNA4 is out in less than 6 months! And we've got info on RDNA5 already:

DGF: A Dense, Hardware-Friendly Geometry Format for Lossily Compressing Meshlets with Arbitrary Topologies (hardware compressed mesh format that's raytracing friendly, already praised by lead designer of Nanite)

"the largest technical leap you will have ever seen in a hardware generation" (PR, but multiple leaks point towards Xbox going w/AMD again)

"RDNA5 to feature entirely chiplet arch" (according to patents)

"Medusa Point Zen 6/RDNA5 in 2026" (Zen 6/RDNA5 already scheduled for laptops/mobile in 2026)

Supposedly targeting late 2025 for desktop versions, though who knows if that date will be hit or if CES - Computex 2026 is more likey. Still, that's a heck of a lot info, implied and leaked (yes Medusa as a codename has come up more than once and looks to be guaranteed as the actual codename). Feels like enough to start a thread.

Other leak points towards some sort of increased AI instruction throughput. Maybe it's that block FP16 (an entire set of matrix multiplication instructions share the same exponent) that AMD is so fond of (already shipping in XDNA2, confirmed support in MI350)

Both Microsoft and Sony are likely to go AMD again for the next gen consoles due to backwards compatibility reasons. In terms of technical leaps, I would say the current gen consoles with the move to a Zen based CPU instead of low power, RDNA 2 graphics and SSD storage brought about the biggest technical leap till date. Not really sure what further leap is possible with the exception of a huge increase in ray tracing performance.

The Medusa APU name has leaked before and yes rumoured to be RDNA 5 based, skipping RDNA 4 on mobile altogether.

And RDNA5 is supposed to bring high end/halo parts like RDNA 2/3 did. Though late 2025 seems a bit optimistic given that they haven't even released RDNA 4 yet. There were rumours that the higher end RDNA 4 parts were cancelled to accelerate RDNA 5 instead but even then I wouldn't expect them before 2026.

 

[pTmdfx]

MI325 (GFX940): 24Q4
MI350 (GFX950): 25Q4
MI400 (???): 26Q4

RDNA4 (GFX12): 25Q1
“Navi 5x” (???): 26Q4 allegedly

Hot news from the rumour mill says MI400 and “Navi 5x” would be the first “UDNA” products.

If “Navi 5x” is indeed a late 2026 product, the timeline does seem plausible. If true, that implies the “unification” decision was likely taken a couple of years ago, since it is now (Nov 2024) ~2 years away from the alleged launch date for both MI400 and “Navi 5”. Supposedly at this point, the SoC/chip lineup should have been decided, and the SoC physical designs should have commenced.

-

While said hot news also suggested this to be “GCN-based”, I find it hard to picture AMD architects wanting to ride on the GCN-classic (Wave64 on SIMD16 over 4 cycles) pipeline for yet another 5 years. That is leaving all the ILP and lower execution latency on the table.

So it is nice to hear a similar opinion in that UDNA is probably gonna be a RDNA descendants/look-alike with CDNA features bolted on:

https://twitter.com/x/status/1859879410196955434

 

[del42sa]

If “Navi 5x” is indeed a late 2026 product, the timeline does seem plausible. If true, that implies the “unification” decision was likely taken a couple of years ago, since it is now (Nov 2024) ~2 years away from the alleged launch date for both MI400 and “Navi 5”. Supposedly at this point, the SoC/chip lineup should have been decided, and the SoC physical designs should have commenced.

it´s pretty hard to believe in such story , because that would make whole RDNA family GPU roadmap pointless.... It´s more likely very recent decision.

While said hot news also suggested this to be “GCN-based”, I find it hard to picture AMD architects wanting to ride on the GCN-classic (Wave64 on SIMD16 over 4 cycles) pipeline for yet another 5 years. That is leaving all the ILP and lower execution latency on the table. So it is nice to hear a similar opinion in that UDNA is probably gonna be a RDNA descendants/look-alike with CDNA features bolted on:

I agree. This is most plausible scenario, they using RDNA5 as a base of the architeture whille making additional changes to fully support all CDNA features and call it UDNA. This explain, how they could get UDNA out so quickly, as usually making new GPU arch takes up to 4 years to develop.

 

[DegustatoR]

While said hot news also suggested this to be “GCN-based”, I find it hard to picture AMD architects wanting to ride on the GCN-classic (Wave64 on SIMD16 over 4 cycles) pipeline for yet another 5 years. That is leaving all the ILP and lower execution latency on the table.

If the idea is to push as many flops as possible and use that for graphics (expecting that graphics is moving onto compute more and more each cycle) then moving back onto GCN approach may in fact be beneficial even despite the relative issues it has. GCN h/w in gaming was never really constrained by its compute performance, all the issues stemmed from its purely graphical frontends/backends. If these would be unnecessary then GCN would likely do just fine. Even if these would just be adequate it would do fine -- although I wonder if it would still need to rely heavily on async compute to fill in state change bubbles whenever these would be occuring.

As for the timelines I'd say chances are high that the whole idea came along with the cancellation of RDNA4 high end - and probably while looking at what they can provide to Sony/MS for the next console gen h/w. So in this regard it does seem more plausible that "UDNA" is a "cancellation of RDNA5 and using CDNA3 to make the next gaming lineup".

 

[Kaotik]

it´s pretty hard to believe in such story , because that would make whole RDNA family GPU roadmap pointless.... It´s more likely very recent decision.

AMD has never shown RDNA roadmap past RDNA4

 

[Dangerman]

Since RDNA 5 is pretty much canned wouldn't this thread be better renamed to UDNA?