The AMD RDNA4 architecture and the future Navi 41, Navi 42, and Navi 43 GPUs have been extensively covered by RedGamingTech. This is one of the first leaks from a known leaker, and while RGT has a track record of reliability, this article is still being treated as a rumour at this time because the information this early is often in flux and subject to change. So, salt shakers at the ready, let’s dive into this incredible delight of a leak from RGT.
The first item of business is that RGT confirms that the ALU count for RDNA4 CUs will stay exactly the same as RDNA 3 – that is, 128 ALUs per CU. GDDR7 memory is undoubtedly in play but has yet to be confirmed, and PCI-E 5.0 is extremely likely to be the standard of choice over PCI-E 6.0 because the latter will most likely not be ready in time.
Interestingly, the RDNA4 architecture will seek a raised 3.5 GHz clock frequency, which when paired with the other architectural changes represents a net performance per watt gain of roughly 50-60%.
One of the shortcomings of AMD’s RDNA3 architecture was the lack of ray tracing and AI/ML performance, which AMD hopes to address in RDNA4. Caches will be significantly upgraded, with the introduction of the 3rd Generation Infinity Cache and a new data prefetch architecture.
WMMA (Wave Matrix Multiply-Accumulate) V2 will also allow for two times the speed per CU for (supposedly) matrix instructions, allowing for full SIMD lane use. Despite the fact that the total increase in CUs is roughly 50%, the performance gain from generation to generation is 100%, which is fantastic to hear.
AMD Navi 41 – 144 CUs
However, the Navi 41 is said to have up to 32 GB of VRAM, most likely of the GDDR7 sort. A second version with 24 or 48 GB of GDDR7 and 6 MCDs is also available (the 32 GB variant has 4 MCDs). The Navi 41 will contain 144 CUs, which when clocked at 3.5 GHz and divided by 128 ALU per CU should result in a whopping 129 TFLOPs of FP32 performance. That is, without a doubt, a Godzilla level of performance.
AMD Navi 42 – 96 CUs
Navi 42 will be slightly muted, with a total of 12,288 CUs clocked at 3.5 GHz, yielding about 86 TFLOPs of FP32 performance. 96 CU is also the exact count of the RDNA3 flagship, the Radeon 7900 XTX, hence this SKU will represent the generation-over-generation comparison between RDNA3 and RDNA4 (or more accurately, as apples to apples its ever going to get). Specifications like this can and do change before the final edition, as RGT acknowledges in their slides.
AMD Navi 43 – 48 CUs
Finally, there will be the Navi 43, which will have 48 CUs. This will produce roughly 43 TFLOPs of FP32 performance when clocked at around 3.5 GHz. This is most likely the lower end of AMD’s RDNA4 portfolio, and it only goes to show how much performance is packed into this new RDNA4 design, where the “lower end” easily exceeds 40 TFLOPs. The memory standard will remain GDDR7, and it will include two MCDs. This is also the only SKU where RGT mentions that the GPU may not be of an MCM design, despite being promised that it is.
Also Read:
