AMD ‘Navi4C’ Based RX 8900 XTX Design Surfaces

Moore’s Law Is Dead has shared an image of what would have been the RDNA4 flagship. For the unaware, AMD has canceled all plans for Navi41 and Navi42. You can read more about that here.

Navi4C Diagram Surfaces

The monstrous package features 3 different dies, the AIDs, SEDs and one MID. This is huge as compared to RDNA3. The diagram only represents the side view of the alleged Navi4C chip, so expect an even more ambiguous design if we get a better overview.

Over a single package lie at least 3 Active Interposers highlighting a 3D architecture. Furthermore, each interposer packs 3 SEDs (Shader Engine Dies) for up to 9 SEDs. To the left of the structure, we see a single MID (Multimedia and I/O Die).

MLID mentions that there probably are more Infinity Cache/Memory Controller dies, obstructed by the limited viewing angle. RDNA4 was moving from 7 chiplets on RDNA3 to 13-20 chiplets in just one generation. The added complexity would explain better why AMD decided to cancel high-end RDNA4.

Basically, AMD divided RDNA3 even more, introducing more and more scalability. Although this design shows had RDNA4 even reached end-users, it would have taken a miracle to do so.

To back this up, we have a patent that shows a design similar to what MLID leaked. Here multiple Virtual Compute Dies are interconnected via a single Bridge Chip, showing a close resemblance to the COW-L Silicon Bridge in the above diagram.

What Remains of RDNA4

Based on the information we have, only Navi43 and Navi44 are what remain of RDNA4. We could see budget Navi44 GPUs using a 6nm process arrive sometime in mid-2024. Similarly, AMD may opt for a 3nm node for Navi43, allowing more CUs in a smaller package.

As ‘Blackwell launches in 2025, AMD should have ample time to launch RDNA5. The current estimate is that RDNA5 may launch 5-6 months after Blackwell, but that’s beneficial if it creates a situation similar to Ampere vs RDNA2.

Source: MLID