The "Navi 4C" GPU from AMD has been described in detail, revealing that the Shader Engines are separate chiplets.

"Navi 4C" is an upcoming high-end GPU from AMD that is unlikely to be released due to the company's shift away from the high-end GPU market in its next RDNA4 generation. In order for AMD to continue investing in the development of this GPU, there would need to be better sales in the gaming graphics card segment, particularly in the high-end, which has not been the case. Moore's Law is Dead has obtained details about AMD's plans for a highly disaggregated GPU, which could have been a fascinating technological endeavor.

AMD's current "Navi 31" GPU features a disaggregation of the main logic components, with the latest 5 nm foundry node located in a central Graphics Compute Die (GCD), surrounded by up to six smaller chiplets built on the older 6 nm foundry node. These chiplets contain segments of the GPU's Infinity Cache memory and memory interface, hence the name memory cache die (MCD). With "Navi 4C," AMD intended to further disaggregate the GPU by identifying additional components on the GCD that could be spun out into chiplets. They also planned to break up the shader engines into smaller self-contained chiplets, which would result in greater yields and lower foundry costs. AMD would have used various packaging innovations to ensure efficient communication between the chiplets.

If AMD had continued to use GDDR6 memory instead of the newer GDDR7 standard, they would have likely retained the 6 nm MCDs from the current generation to provide video memory interface and last-level cache to the GPU. This would minimize R&D costs and take advantage of the reduced foundry costs for the 6 nm node.

AMD identified the GPU's media acceleration engine and Radiance Display Engine as components that could be disaggregated in the next round. These fixed function hardware components could potentially make use of older foundry nodes. They would be spun off into a separate chiplet called MID (media and I/O die). The specific foundry node for the MID is unknown, but considering that AMD is able to incorporate the latest media and display I/O features into the 6 nm "Navi 33" monolithic silicon, it is possible that they would use the older node.

The most critical part of the GPU, the Shader Engines, would be spun out into chiplets called SEDs (shader engine dies). Each Shader Engine consists of a fixed number of workgroup processors (WGPs). These chiplets would be built on an advanced foundry node, potentially the same one used by NVIDIA for its next-gen "Blackwell" GPUs.

The SEDs would be seated on active interposer dies (AIDs). An interposer is a silicon die that facilitates high-density wiring between stacked chiplets. The "active" part of the AID refers to its ability to facilitate wiring among dies stacked on top and to neighboring AIDs. TSMC has developed COW-L (chip-on-wafer-L) bridges for inter-AID high-density wiring.

Communication between the AIDs and the MID is still unknown. The current generation of MCDs are connected to the GCD using Infinity Fan-out Links, a high-density wiring method that utilizes the fiberglass substrate. If AMD were to use a different method to connect the MCDs, it would indicate the use of a newer generation. In addition to COW-L bridges, AMD is also utilizing TSMC's COW-V TSVs (through-silicon via) innovations to connect the SEDs to the package substrate.

Unfortunately, it is highly unlikely that "Navi 4C" will ever be developed. AMD has already implemented many of these packaging innovations in its latest MI300 compute processor, based on the CDNA3 architecture. It would have been incredible to see these innovations in the gaming graphics segment, but economic factors prevent AMD from investing further in the development of "Navi 4C." The gaming graphics card market is currently experiencing a significant downturn, and the enthusiast-class GPU market is niche.

The current market conditions are very different from 2021, when the crypto-currency mining boom drove demand for larger GPUs. AMD is rumored to have launched its RX 6950 XT GPU towards the end of the crypto boom, which may have been mistimed. By that point, "Navi 31" was already well-developed and ready for mass production. Unless there is another dramatic surge in demand for high-end GPUs, it is unlikely that AMD will justify the cost of developing "Navi 4C" beyond the concepts discussed in this article.