Spectacular AMD Navi 31 RDNA 3 MCM specs indicate 15,360 SPs, 32 GB 256-bit GDDR6, 2.5 GHz boost, and 75 TFLOPs, Nvidia Lovelace AD102 could go up to 92 TFLOPs

2022 is poised to be an exciting year for GPUs — Intel is all set to debut its mainstream Arc Alchemist offering while significant iterative upgrades are on the anvil for AMD Radeon and Nvidia GeForce cards. While we've seen some early rumors about Navi 31 and Lovelace before, some additional information has just trickled in.

Once again, @greymon55 on Twitter seems to have gathered a few details about Navi 31 and Nvidia AD102 GPUs. We have known so far that Navi 31 RX 7900 XT (?) will be a multi-chip module (MCM) configuration featuring a combination of 5 nm and 6 nm parts with 256-bit GDDR6 memory. @greymon55 now notes that Navi 31 will feature 120 workgroup processors (15,360 streaming processors) and a 256-bit 32 GB GDDR6 memory that may run at 18 Gbps.

For perspective, the Radeon RX 6900 XT features a Navi 21 XTX GPU with 5,120 SPs and 256-bit 16 GB GDDR6 VRAM at 16 Gbps, so we are looking at 3x the shading unit count and twice the VRAM with Navi 31.

Navi 31 is also likely to offer 256 MB or 512 MB of Infinity Cache. Information on the exact amount of Infinity Cache is still scarce and not confirmed. While most sources point to a 256 MB cache, RedGamingTech says a 512 MB cache is not out of bounds either. What they all seem to be sure of is that RDNA 3's Inifinty Cache will be a 3D stacking solution unlike what we've seen with RDNA 2. A clock speed of 2.4 to 2.5 GHz is likely yielding an FP32 performance of 75 TFLOPs.

Coming to Nvidia Lovelace AD102, we already know that this will be a 5 nm GPU with 18,432 CUDA cores in 144 streaming multiprocessors. AD102 will offer 24 GB GDDR6X memory on a 384-bit bus running at 21 Gbps. RedGamingTech notes that GDDR7 is also being tested by Nvidia, but that's for a future architecture and not Lovelace.

Interestingly, @greymon55 says the FP32 performance could be between 85 to 92 TFLOPs, which is a significant bump over the 66 TFLOPs we heard previously. Of course, TFLOPs alone are not an indicator of performance but if true we stand to see about a 1.4x increase from Ampere. Nevertheless, at least a 2x to 2.4x performance increase in general can be expected in moving from Ampere to Lovelace as per current rumors.

It is still quite early to speculate — the cards are only expected sometime in Q3 2022 anyways — but it looks like AMD Navi 31 may end up costing more than Nvidia AD102 given the seemingly complex chip architecture. However, AMD will continue to use the cheaper GDDR5 VRAM and rely on advances in 3D Infinity Cache in order to remain competitive in terms of pricing against Nvidia. 

Buy the EVGA GeForce RTX 3080 Ti FTW3 on Amazon