The truth is that AMD advanced it at the time and it was really hard to believe, because it was something more informal, like a supposed advance that would be possible in the future and that it could happen or not. But at FDA 2022 last night, the company laid the foundations for what the new RX 7000 GPUs will be and as such, part of the rumors have been confirmed and some details are surprising. So will be the AMD RDNA 3 GPU.
Confirmation here and there, rumors that are right and others that are not, but in the end the riverbed narrows. What was revealed yesterday by AMD is really little, very brief, but it leaves great impressions and invites us to delude ourselves about a possible surprise to NVIDIA.
AMD RDNA 3 with RX 7000, small big improvements
We’ll start by saying that the main thing is the announced 50% increase in performance per watt. It’s the same jump from RDNA to RDNA 2 and now it’s happening again in RDNA 3, which is very impressive considering that the node used is the 5nm from TSMC. So how has AMD done it? Well, there are quite a few important changes, which are small in individual terms, but which together represent large improvements. (herbsforever.com)
Starting with the fact that AMD will use chiplets in this architecture, but not as expected apparently. Everything revolves around changes in Compute Units or CUssince they have been redesigned within the new architecture with an “alleged” premise: there will be fewer of them, which is an interesting change of concept.
It looks like there will be more WGP, but instead they will have dual CUs with double the SIMD32. This is interesting, because it is a necessary step for the move to chiplets that we have mentioned. And it is that this step is the most important, because AMD would have left a chip with the WGP and L0 and L1 as the main one, while it would have removed the L2 and the Infinity Cache to several chiplets mounted on an interposer and connected by Infinity Fabric .
Five chiplets at its maximum configuration
And it is that AMD is going to surprise with an unusual configuration and that is totally revolutionary, since from now on we have to talk about GCD or Graphics Compute Dies (GPU cores as such) and MCD or Multi-Cache Dies (the annexes where they take out the Infinity Cache, and supposedly the L2).
What we do not know because it has not been specified as such is in what lithographic process each one arrives. It is speculated that TMSC will manufacture the GCD at 5nm, while the MCDs could come from Samsung or the Taiwanese.
Going back to the architecture, each GCD is equipped with 4 Shader Engines and each of them will include 2 Shader Arrays, where at the same time each of them has 4 WGP, making a total of 32 SMID32 ALUs, giving a total count of 8192 Shaders on the midrange chip.
The oldest of them is going to 48 WGP with 12 SA and 6 SE to do 12,288 coreswhile Navi 33 will have 5120 total. This is done this way because it is estimated that AMD could include Navi 32 instead of Navi 31 in the market first, leaving NVIDIA in the high range and focusing on the medium to capture more sales.
The dates are unknown at the moment, like the prices, but it is rumored that they could arrive between October and November of this year.
