I’ve been thinking of when will the RDNA4 cards come out.
As MILD mentioned RDNA 4 will come out around Q3 2024. I don’t think there will be RDNA 3.5 refresh cards or RDNA 3 refresh cards out next year.
I think RDNA4 will be very similar to RDNA3 apart from very small arch improvements and update to Raytracing core.
There were rumors in 2022 that AMD had issues with TSMC 3nm node and that they will be using 4nm. Current rumors don’t say anything about which node will RDNA4 use but, seeing that TSMC N3E 3nm node is just being put in production and others will be using it, makes sense that AMD has to use the 4nm node in 2024. This could make AMD release the RDNA 4 before Nvidia does its new series. So, I’m thinking RDNA4 could come out end of May and be available in June 2024.
Further, I was looking at how big the RDNA 4 flagship chip will be in mm2 and what its performance could be. Taking the N31 which is based on 5nm and 6nm nodes and combined size of 530mm2. An RDNA4 best would be around 370mm to 450mm2 chip with 90-96 CUs like Rx7900 series, but with 256Bit bus, faster memory since it will use Gddr7 and rated TDP of below 280W. I came to this conclusion that 4nm TSMC node is a very small improvement in transistor density, of just 6% for the N4, (N4x or Nvidia specific 4N might be a bit more).
Looking at the 4nm node and doing the math is no wonder that AMD can’t produce a high-end GPU next year because by my math comes out that a 20-30% more performat GPU then a RX7900xtx would have to be bigger then 680mm2 and have a TDP of 410W, that’s what the 4nm node does.
But here are the all the good things, the GPU, let’s say it’s called Rx 8800 XT is out in middle of next year has 16gb of Vram for 600$, identical performance in raster compared to Rx 7900 XTX and somewhat better performance in raytracing.
There are two AMD patents on raytracing that I’ve read few months back. The first one, released 1 year prior to first RDNA 2 GPU, talks about raytracing core. But the second one, it was released in June this year. So, the latest AMD patent describes GPU withing its raytracing core, addition of a hardware specific traversal engine and specific BVH memory cache. Not to go into details, from what I understand of the two patents first patent describes raytracing core in RDNA 2&3 and the second patent describes similar but a much improved way of doing raytracing. I’m hopeful we will see this is RDNA4(the patent did arrive this june and next june we’ll have RDNA 4 card so it matches the schedule prior to RDNA2) (https://www.freepatentsonline.com/20230206543.pdf)
I see no one, coming to the conclusion that there’s only leaks about N44/48 because AMD doesn’t need a larger die for a multiple GCD (gpu chiplet MCM approach) SKU.
Put two N44s together, and you’ve got like 80% of the equivelant of 2x 7900xtx…quite capable or competing with a 5090 in theory.
Bcuz it sounds like copium lol
Except scaling never works like that, esp not with adding in the problems of chiplets
Erm
It was a spit ball % guess…chiplets for CPU scale excellently (up till about 64 cores, sort of drops off steeply after that)…whether GPUs scale like that remains to be seen (doubtful)…but if you think AMD hasn’t gone to all this trouble to break away from monolithic designs with MCDs and GCDs, and not iterate with a multi GCD design…then I dunno what to tell you bro…🤷
It’s not as impressive as you make it out to be. Splitting the MCDs and GCDs is certainly pretty nice, but both Intel and AMD have shown to have better and more advanced packaging capabilities in their GPUs- with MI300 and PVC- the only reason they haven’t come to consumers yet is cost and complexity chiefly.
However, if AMD using something MI300esque with RDNA 4… and failed, then yes, it stands to reason that only the monolithic skus would remain.
Alternatively, the base RDNA 4 arch could just be so cooked they thought it wasn’t worth the effort of developing the more expensive and complicated chiplet skus.
Or who knows, maybe it’s a combination of the two, or something else.
Also, the idea that AMD has N44/N48 and can just glue the two together to act as their flagship is also wrong. There has to be additional interconnect logic among other things added to the two dies. If the chiplet dies are canned, then they have to do expensive and time consuming respins on their existing planned RDNA 4 dies (N44/48) in order for them to be allowed to be used in chiplet designs.
There’s nothing to say, that isn’t the case (with what you suggested in the last part of your message)…who knows? All these leakers throw so much shit at the wall and tiny bits of it stick, and they point it out and jump up and down, and say “I told you so!” Even though there’s been a tonne of false leaks and misdirection that everyone has had to sift through along the way.
Take for instance the recent leak of Linux driver updates, mentioning gfx1200, could be a larger “N41” die, or it could be a nothing-burger.
yeah I think that’s clearly going to be what they do whenever they decide to do chiplets again