AMD Navi 41, Navi 42 and Navi 43 GPU get rumored preliminary specifications for the upcoming Radeon RX 8000 series
In the first order of business, RGT confirms that the ALU count for the RDNA4 CUs will remain exactly the same as RDNA 3 – which is to say it will remain at 128 ALUs per CU. GDDR7 memory is definitely in play but not confirmed at the moment although PCIE 5.0 is very likely to be the standard of choice versus PCIE 6.0 as the latter will likely not be ready in time. Interestingly, RDNA4 architecture is going to be targeting a boosted 3.5 GHz clock frequency and this combined with the other architectural movements means you are looking at a net performance per watt gain of around 50-60%.
One of the things that was lacking with AMD’s RDNA3 architecture was the ray tracing and AI/ML performance and AMD is looking to fix that with RDNA4. Caches will get big upgrades and 3rd Generation Infinity Cache is going to be rolled out along with a new data prefetch system. WMMA (Wave Matrix Multiply-Accumulate) V2 will also enable 2x the performance per CU for (presumably) matrix instructions allowing full SIMD lane usage. Even though the total increase in CUs is around 50%, the performance increase generation over generation is 100% - which is superb to hear. Given below are the separate dies:
AMD Navi 41 – 144 CUs
Here is where things get even crazier, Navi 41 is touted to have up to a massive 32 GB of vRAM – probably of the GDDR7 variety. A second configuration also exists with 24 or 48 GB of GDDR7 and 6 MCDs (the 32 GB variant has 4 MCDs). The Navi 41 will have 144 CUs, which clocked at 3.5 GHz and at a calculation of 128 ALU per CU should yield a whooping 129 TFLOPs of FP32 performance. Needless to say that is a Godzilla level of performance right there.
AMD Navi 42 – 96 CUs
Navi 42 will be slightly muted and have a net total of 12,288 CUs which clocked at 3.5 GHz should yield around 86 TFLOPs of FP32 performance. 96 CU is also exactly the count of the RDNA3 flagship, the Radeon 7900 XTX, so this SKU will represent the apples to apples generation over generation comparison between RDNA3 and RDNA4 (or more accurately, as apples to apples its ever going to get). As RGT notes in their slides, specifications like these can and do change before the final revision.
AMD Navi 43 – 48 CUs
Finally, we have Navi 43, which is going to have 48 CUs. Clocked at around 3.5 GHz, this will yield around 43 TFLOPs of FP32 performance. This will probably be the lower end of AMD’s RDNA4 lineup and just goes on to show how much performance is being packaged in this upcoming RDNA4 architecture where the “lower end” easily exceeds 40 TFLOPs. The memory standard is still going to be GDDR7 and it is going to come with 2x MCDs. This is also the only SKU where RGT notes that the GPU might not be of an MCM design – although they are being told this is currently the case.
