AMD Radeon 7000/RDNA3 Thread

Edit: Reviews starts at around page 53 - https://forums.overclockers.com.au/threads/amd-radeon-7000-rdna3-thread.1296854/page-53



Hype?

 

Does it say when?

 

You'd assume late 2022

 

I think earlier than this. macOS 11 GPU drivers already contain references to Navi 31. And that was 6 months ago - https://videocardz.com/newz/amd-navi-31-spotted-in-apple-macos-11-big-sur-operating-system

 

After reading a bit more you might be right, seems like RDNA 3 might be more like RDNA2+ though

 

How so? It's going to be an MCM design utilising infinity fabric and some advanced logic to direct workflow. That's a pretty massive change.

 

It is a big change in terms of raster performance, and architecture, and builds a great solution for future generations with additional chiplets being added.
I wonder if raytracing will get IPC improvements also, or are they going to solve that with the brute horsepower method from the doubling of grunt/cores/CU's

 

So the 80 CU MCM design rumour is being walked back by the originator of the rumour. It start with this tweet of his earlier in the month. Now over at the Anandtech forums the rumour poster has clarified his rumour;



Source: https://forums.anandtech.com/thread...a2-architectures-thread.2589999/post-40420465

He does has some internal access as a couple of days later after his initial MCM tweet, an AMD MCM patent for GPU's was published;

https://www.freepatentsonline.com/y2020/0409859.html

3dcenter has more on this patent - https://www.3dcenter.org/news/amd-arbeitet-am-technischen-unterbau-von-multichip-grafikloesungen (google translate comes on automatically and it reads nicely).

Other rumours currently surrounding RDNA 3;

• a 2.5X6900XT performance target from RTG as above,
  
• CapFrameX claims his source is promising 32GB vram for the top RDNA 3 sku. (tweet)

If AMD is using MCM for RDNA3, I think it makes more sense to make 40CU chiplets and build up the medium and highend sku's from these. Part of the charm of a chiplet strategy is build cheaper small dies and bin them accordingly.

 

yeah that makes a lot of sense to use much smaller chiplets. also aligns with the piece I can't remember where said there would be a master chiplet that communicates with CPU, all the other chiplets only talk with the master

 

Thermally/electrically 2 x 80 won't work. Thats a 500-600w card. At standard clocks, volts, and even undervolting that is going to be a hard ask. Also your memory issues at 1440p and particularly at 4k are no better.

AMD could certainly make something around a 60 CU die, with the top product being 2 x 60 CU. 384 bit bus with cache would provide sufficient performance.
AMD Could then have a 40-60 CU 192 bit bus

With a 60 CU die, you can harvest 50, 40, 30, etc from that die. Two perfect dies would be rare. Regular high end would be say 2 x 50 (7800xt). Low high end would be (7800) would be 2 x 40.

 

I won't rule out a 160CU sku at this stage. AMD's cache patents from last year have some crazy efficiency gains. I would not be surprised to see greater than +50% performance/watt gain for RDNA3 over RDNA2 without having to go to TSMC's N5 process.

TSMC's yields are very very good atm. Perfect dies wouldn't be that rare. Perfect dies that can run like the clappers might be rarer yet. But if you were going for a 120 CU sku, you may as well go for 40 CU chiplets and put 3 dies into the MCM design. AMD's MCM patent says that their idea can support 4 dies and can go higher with different layouts.

Smaller dies are better if you are already in the MCM-land. Dies become cheaper and the binning process becomes more accurate. And those perfect run-fast dies become more common.

They certainly wouldn't harvest 30CU from 60CU dies. 50% wastage from precious silicon is just not needed given TSMC's yields.

 

Its not just yields its quantities.

160 CU is going to gobble up pretty much all of the volume. You will sell half as many cards, so you will have to sell them for twice as much just to make the same money. Plus any glue silicon would mean more than that. Given you can sell a 80 CU for $2k that doesn't seem to be a sensible decision. Selling $4k or $5 cards, while possible, will only net them back in the same position they are now.

AMD would be ruling the roost if they had 120 - 100 - 80 - 60 - 40 stack. 120 and 100 would still be much faster than the previous gen. 6800xt and 6900xt are limited by memory bandwith more than anything else, so even a minor improvement per CU would see big gains at say 4k.

AMD and Nvidia need to work out a way they can make more cards.

 

Well the roadmap shows them being on 5nm, so it'll depend more on the transistor density. If an 80CU die on 5nm winds up being significantly smaller than the current 7nm ones, they could easily sell the 2x80CU cards for USD1500 and simply smash nVidia. That'd only be a 50% price increase (matching the 3090 RRP) and it would destroy everyhing performance.

 

Just on this. All AMD has said is 'Advanced Node'.



This could be one of TSMC's N7p (0 EUV layers) or N6 (4 EUV layers). It might not mean their N5 process (10+ EUV layers). Apple is using loads of TSMC's N5 capacity atm, and its not really until they step off it to go to N3 (20+ EUV layers) that there will be a large capacity of N5 available for everyone else (depending on just how many EUV machines TSMC gets since Intel and Samsung also wants them now). This is H2 2022 when Zen 4 on N5 is already tentatively mapped out to appear.

I honestly can't see AMD waiting to late 2022 to release RDNA3. They'll have something inside 12 months. And I suspect it won't be on N5 but on N7p since EUV machines are golden atm and they will be reserved for serverside sku's (or custom Apple sku's) where premium means dollars.

 

Right, must have been confusing it with the Zen 4 slide.

 

hmm... in the roadmap above is it wrong to interpret the dots on the horizontal axis as being the beginning of the respective year? That would show that AMD is aiming for RDNA3 as being released before the beginning of 2022...

Just a Friday alcoholic thought

 

MLID has another leak video on RDNA3, nothing amazing but it's something:
Some quick notes of the full of salt predictions:
1. Generally it's too early get anything concrete, again 3x 6900xt performance is bs. Debunks a bs die shot floating around.
2. Will have At least 1 io die and at least 1 compute die. Having a single IO die essential to appearing to OS as a single GPU.
3. earliest release early 2022 but probably mid 2022. Nvidia will likely have an ampere refresh to semi counter it (but RDNA3 should come out on top).
3. expects 60-80% improved performance and AMD has publicly stated already aiming for at least 50% perf/watt improvements over RDNA2
4. Ray tracing and geometry per CU improved

 

From earlier in the month. I completely missed this one...


https://twitter.com/KittyYYuko/status/1388904408394571784

This seems to suggest that Navi 33 will have 80CU's. Which sounds far fetched. But they are a reliable leaker apparently.

And then this...

https://twitter.com/Broly_X1/status/1398912465270755331

 

Yeah seem a bit far fetched. I'd think more like Navi 32 = 80cu, double navi 22 which given 5nm 1.8x density gains is plausible.

 

so mcu/mcd chiplets next year? Wozzle i cant wait?!?!?