Discussion RDNA4 + CDNA3 Architectures Thread

[DisEnchantment]

With the GFX940 patches in full swing since first week of March, it is looking like MI300 is not far in the distant future!
Usually AMD takes around 3Qs to get the support in LLVM and amdgpu. Lately, since RDNA2 the window they push to add support for new devices is much reduced to prevent leaks.
But looking at the flurry of code in LLVM, it is a lot of commits. Maybe because US Govt is starting to prepare the SW environment for El Capitan (Maybe to avoid slow bring up situation like Frontier for example)

See here for the GFX940 specific commits

History for llvm/lib/Target/AMDGPU - llvm/llvm-project

The LLVM Project is a collection of modular and reusable compiler and toolchain technologies. - History for llvm/lib/Target/AMDGPU - llvm/llvm-project

github.com

Or Phoronix

More AMD "GFX940" Enablement Work Landing In LLVM - Phoronix

www.phoronix.com

There is a lot more if you know whom to follow in LLVM review chains (before getting merged to github), but I am not going to link AMD employees.

I am starting to think MI300 will launch around the same time like Hopper probably only a couple of months later!
Although I believe Hopper had problems not having a host CPU capable of doing PCIe 5 in the very near future therefore it might have gotten pushed back a bit until SPR and Genoa arrives later in 2022.
If PVC slips again I believe MI300 could launch before it

This is nuts, MI100/200/300 cadence is impressive.



Previous thread on CDNA2 and RDNA3 here

Question - Speculation: RDNA3 + CDNA2 Architectures Thread

Man I have been dying to make this one for a while now. First rumours for RDNA3 are here so new thread time! Just going to start off with this one for now: kopite7kimi on Twitter: "@VideoCardz Ah, I mean a simple mcm design with 10240 cores is not enough. Because the lift from RDNA2 to RDNA3...

forums.anandtech.com

 

[adroc_thurston]

Rdna1 ->RDNA2 was 1 year 4 months gap not full 2 year cycle

That wasn't a new uArch.

 

[gaav87]

It's not.

Summer 2026 is more likely like rdna1

 

[gaav87]

That wasn't a new uArch.

Well even in the roadmap it shows ~sept so not full 2 year

 

[adroc_thurston]

Well even in the roadmap it shows ~sept so not full 2 year

It's full 2 years.
Maybe more even.
But hey, fresh new IP.

 

[ToTTenTranz]

It's not.

Is next-gen RDNA coming only 2 years after RDNA4? Are they keeping the 24 month cycle between gens?

 

[adroc_thurston]

Is next-gen RDNA coming only 2 years after RDNA4? Are they keeping the 24 month cycle between gens?

Wdym only? 8q is the norm for big boy GFX IP.

 

[Kepler_L2]

Summer 2026 is more likely like rdna1

October 2026 AFAIK.

 

[SolidQ]

3 days left, and zero leaks about perfomance from AMD/NV. Seems both companies hiding well

 

[ToTTenTranz]

Wdym only? 8q is the norm for big boy GFX IP.

Initial rumors about high-end RDNA4 being cancelled were also telling this was decided because AMD was fast-tracking RDNA5.

Big Polaris was cancelled at some point, but Vega appeared 14 months later.

 

[beginner99]

The buying decision is quite simple when one buys a ~$1000 consumer dGPU - the consumer wants the full feature set with the "standard experience". That's always been GeForce.

I can get behind the decision not to compete in the high end. But just offering mid-range at $25-$50 lower with general worse software is not really competing or challenging the competition.

 

[jpiniero]

I can get behind the decision not to compete in the high end. But just offering mid-range at $25-$50 lower with general worse software is not really competing or challenging the competition.

Well, the 5070 is expected to be $699.

If AMD is really doing two 16 GB models you figure the price gap between the two won't be much. Maybe if you are lucky it'd be $549/$599.

 

[Win2012R2]

3 days left, and zero leaks about perfomance from AMD/NV. Seems both companies hiding well

It's been very tight this time around! Amazing work by counter-intel guys at NVidia and at AMD they just had nothing to leak about...

Well, the 5070 is expected to be $699.

If 5080 is indeed again $1200+ then 5070 seems unlikely to be $699...

 

[yuri69]

Initial rumors about high-end RDNA4 being cancelled were also telling this was decided because AMD was fast-tracking RDNA5.

Big Polaris was cancelled at some point, but Vega appeared 14 months later.

Fast-tracking anything is... not a thing easily done in the chip design word, moreover in AMD's mindset.

As for the Big Polaris, there were 3 chips initially planned in that 14nm time frame - Baffin, Ellesmere, and Greenland. They become Polaris 10, Polaris 11, and... Vega 10. The HBM2 Greenland was even supposed to be used in a Zen 1 APU.

Most probably there was no fast-tracking with the GCN 5 gen, since this all-in-HBM2 Vega gen was more like cancelled:
* Vega 12 sold in low quantity as an Apple exclusive and more or less replaced by Polaris 21
* Vega 11 cancelled completely and replaced by Polaris 20
* Vega 10 made it to consumers

---

Btw that roadmap shows the number of chips going down. With shrinking market share and rising development expenses, it would IMO make sense to keep it to 1-2 AMD dGPU chips.
* Navi 1 - 10, 12, 14
* Navi 2 - 21, 22, 23, 24
* Navi 3 - 31, 32, 33
* Navi 4 - 48, 44
* Navi 5 - ?

 

[Tuna-Fish]

If AMD is really doing two 16 GB models you figure the price gap between the two won't be much. Maybe if you are lucky it'd be $549/$599.

The cost difference between 20Gbps and 18Gbps GDDR6 is abnormally high for such a small speed bump. This might be reflected in the cards.

If 5080 is indeed again $1200+ then 5070 seems unlikely to be $699...

5070 is the highest-end model that has some competition. I would expect there to be a significant gap in price/perf between it and the next model up (which, according to rumors, will be 5070 Ti).

If the rumors about early 24Gbit GDDR7 modules are true, the 5070 is by far the model that needs them the most. It's just much more viable with 18GB than it is with 12.

Btw that roadmap shows the number of chips going down. With shrinking market share and rising development expenses, it would IMO make sense to keep it to 1-2 AMD dGPU chips.
* Navi 5 - ?

The roadmap shows Navi 5 extending both above and and below Navi 4 in the range, unless they figure out chiplets that don't suck, that probably means 3 designs.

 

[marees]

If the rumors about early 24Gbit GDDR7 modules are true, the 5070 is by far the model that needs them the most. It's just much more viable with 18GB than it is with 12.

Rumour is that 5070 will be 12gb
I expect 5060 ti to be a cut down 5070 & again 12gb

These two should be the main competition to 9070xt & 9070

5060 would have 128bit bus. That would strongly benefit from the 24gb memory

 

[blckgrffn]

5060 Ti as a cut 5070 is a best case scenario IMO - to me another 5060Ti 16GB with higher density chips and the full glow up on the tiny die seems like a winner to me if I am Nvidia.

It barely matters if AMD "is better" because you still have 16GB and DLSS and all that copium.

 

[adroc_thurston]

Initial rumors about high-end RDNA4 being cancelled were also telling this was decided because AMD was fast-tracking RDNA5.

Who said that even.

I can get behind the decision not to compete in the high end. But just offering mid-range at $25-$50 lower with general worse software is not really competing or challenging the competition.

They really don't want to? The only way to challenge the comp in client dGFX is the Chainsaw.

Btw that roadmap shows the number of chips going down. With shrinking market share and rising development expenses, it would IMO make sense to keep it to 1-2 AMD dGPU chips.
* Navi 1 - 10, 12, 14
* Navi 2 - 21, 22, 23, 24
* Navi 3 - 31, 32, 33
* Navi 4 - 48, 44
* Navi 5 - ?

You're going right back to at least 3 dies with RDNA5. Hope that helps!

 

[marees]

Who said that even.

MLID

 

[Bryo4321]

You're going right back to at least 3 dies with RDNA5. Hope that helps!

small, smaller and smallest?
😂

 

[techjunkie123]

small, smaller and smallest?
😂

Hey it's average sized, bigger than average, and pretty big.

 

[Saylick]

Die size always equates performance.

*Laughs in Battlemage*

 

[gdansk]

*Laughs in Battlemage*

If Intel had built a bigger BM it'd have more performance too. I think that's his point. And as far as I know such a part is still a possibility.

 

[adroc_thurston]

small, smaller and smallest?
😂

Small, medium and a modest chungus.

MLID

Gitmo.

*Laughs in Battlemage*

BMG is the posterboy of "PPA matters"!

 

[JustViewing]

Maybe AMD should bring back Raja to lead the Radeon software side. Raja for all his faults related to hardware, did reasonably well in software side. The were lots of improvements in software during his time. BTW, what has happened to periodic Adrenal releases? It got lost after Rajas exit.

 

[Saylick]

BMG is the posterboy of "PPA matters"!

More like ppA.

BMG is a shower, not a grower.