AMD RDNA 3, a riot of chiplets for the Navi 31 GPU? It could have up to seven

The future range of GPU AMD architecture based RDNA 3according to rumors, it aims to achieve unprecedented performance and efficiency by adopting a design Multi-Chip Module (MCM) where, on the same package, multiple GPUs act in concert as if they were one big chip. Something along the lines of what AMD itself did in the world of accelerators for HPC and AI with the Instinct MI250 and MI250X solutions.

In particular, AMD would have chosen this solution for the chips Ships 32 and Ships 31while he would have opted for the “monolithic” road for Navi 33, a GPU that should give life to the rumored Radeon RX 7700 XT, predicted to offer performance almost in line with the Radeon RX 6900 XT.

According to leaker Greymon55, AMD’s Navi 31 project would consist of seven chiplets. Due chiplet should be i Graphic Complex Die (GCD), which is the heart of the GPU where RDNA 3 stream processors and other computing solutions, such as Ray Accelerators, reside. These units should be made with a production process 5 nanometers.

There would be then quattro Memory Complex Die (MCD) a 6 nanometri to manage the interface to the VRAM memory. It cannot be ruled out that these die also integrate Infinity Cachethe large cache that AMD first implemented with the Radeon RX 6000 series RDNA 2 GPUs. Finally, there should be one die tied to I / Oa chip capable of acting as a trait d’union between the various chiplets and the rest of the system.

An undoubtedly complex project, and further complexity could arise if AMD decided to put MCDs on top of GCDs following when done with the Ryzen 7 5800X3D’s 3D V-Cache technology. Not to know if the company has decided to follow this path right away or if it will do so with future generations.

Certainly one LEGO style design should guarantee high production yields, because TSMC is required to make smaller and less complex chips, which should facilitate smooth production. Additionally, AMD gains more leeway in terms of design, as we have seen in the CPU sector with EPYC and Ryzen.

NVIDIAapparently, it does not seem willing to follow it on this path with the next generation of “Ada Lovalace” graphics chips as can also be understood from the recent presentation of the Hopper architecture for the world of accelerators for datacenters and from leaks related to data theft operated by Lapsus $. Future GeForce RTX 4000s will likely continue to rely on a monolithic GPU, built at 5 nanometers.

As for the technical specifications of the future video cards from AMD, expected in all probability in September and October, the rumors are not unique but for the Navi 31 chip there is a total of 15360 units, far more than the 5120 stream processors of the current top-of-the-range RX 6900 XT. AMD could also match the GPU to 32 GB of GDDR6 memory, doubling the amount compared to the 6900 XT, while maintaining a 256-bit bus by virtue of a large amount of Infinity Cache, even greater than the 128 MB integrated on the current flagships. .