The latest tip from Greymon55 indicates that the specifications of AMD’s Radeon RX 7000 GPUs were updated sometime before tapeout. The higher-end Navi 3x cores got a haircut right before entering mass production, most probably to avoid wastage. With RDNA 3, the Work Group Processors (WGPs), the basic unit of scheduling on Radeon GPUs have been overhauled. We’re now looking at four 32 core SIMDs per Compute Unit for a total of 128 per CU and 256 per WGP. Previously, each WGP consisted of two 64-core CUs or four SIMDs.
Previously, it was thought that the Radeon RX 7900 XT (Navi 31) would feature a total of 15,360 shaders across two identical 5nm dies. As per Greymon55, that figure has now been cut down to 12,288 spread across two 6,144 core chiplets. It’s possible that the 6nm MCDs have been tweaked as well as adjusting memory controllers is much easier than modifying the GPU core. Regardless, it seems like we’re still looking at a seven-die part for Navi 31.
Navi 32 is also going to be a chiplet design with x2 5nm Graphics Compute Dies (GCDs) and 4-6 6nm MCDs. Earlier, it was expected to come with 10,240 cores paired with a 192-bit bus and 12GB of GDDR6 memory. Now, the forecast has been updated to 8,192 shaders across 32 WGPs, 8 Shader Arrays, and four Shader Engines.
Navi 33 is unchanged with 4096 stream processors across 16 WGPs. It’s expected to power the Radeon RX 7700 XT with a performance increase of more than 2x if you include the IPC, compute, and frequency gains. It will come with 64-128MB of L3 cache, a 128-bit bus, and 8GB of GDDR6 memory. The Navi 33 core will run at up to 3GHz, with AIB models potentially being even faster.
Since only the higher-end SKUs have gotten reductions in core counts, this has likely to do with optimal scalability. It’s possible that you’re getting roughly the same levels of performance with a 12,288 core Navi 31 die, perhaps even higher with improved boost clocks, as compared to the full-fat 15,360 core die. The same can be said for Navi 32. Navi 33 remains unchanged as it was already limited with respect to the shader strength, to begin with. (armodexperiment.com)