AMD's next-gen RDNA 3 architecture is still a long way away, it'll be interesting to see how AMD deploys its next-gen GPU architecture as it will be offered in MCD format -- or Multi-Compute Die.
MCD is similar to the MCM (Multi-Chip Module) design that NVIDIA will use with its next-gen Hopper GPU, and the chiplet-based technology AMD uses on its various Zen-based CPUs. The latest info on the next-gen RDNA 3 architecture is that Navi 31 will be an MCD (Multi-Compute Die) design and taping out "soon" according to leaker @KittyYYuko, with Navi 32 possibly taping out in Q2 2022.
- AMD Navi 31: Multi-Compute Die (MCD) tape out in Q4 2021
- AMD Navi 32: Multi-Compute Die (MCD) tape out in Q4 2021
- AMD Navi 33: Monolithic GPU (like normal) tape out in Q1 2022
AMD will be making multiple Navi 3X GPUs going forward, with Navi 31, Navi 32, and Navi 33 all teased. As for performance, we should expect the flagship RDNA 3 GPU in monolithic form -- which should take form with a Navi 33-based GPU -- to be somewhere in the 100%+ increase over the Radeon RX 6900 XT range.
- Read more: AMD's next-gen RDNA 3 + Zen 4 architectures both launching in Q4 2022
- Read more: AMD Navi 33 GPU: next-gen RDNA 3 with 80 CUs + 5120 cores = beast GPU
- Read more: AMD RDNA 3 GPUs could 'take performance crown' from NVIDIA in 2022
- Read more: AMD Navi 31 teased: chiplet GPU design should be revealed this year
Infinity Cache is expected to be greatly improved, with AMD's next-gen Navi 31-based Radeon GPU packing up to 512MB of Infinity Cache... this is a 4x increase over the 128MB of Infinity Cache on the Radeon RX 6800 XT and Radeon RX 6900 XT graphics cards under the RDNA 2 GPU architecture.
But now, let's talk about efficiency.
In some other tweets, new leaker on the scene @Greymon55 tweeted out Ada Lovelace followed by three fire emoticons, while RDNA3 had two fire emoticons next to it. @Rule110 replied, saying: "I expect GPU chiplets to be crazy innefficient. I mean Ponte Vecchio is basically a meme. However like HBM on Vega maybe it's also vital for it to use a massive L3 to even function well. I have no idea what Lovelace is expected to bring up. As usual with NVIDIA".
In reply to that, @Greymon55 said: "No, GPU chiplets is very efficient and does not have much impact on performance and power consumption".