Serial leaker @greymon55 (Twitter) has shared the memory configurations of NVIDIA and AMD’s next-gen graphics cards. In short, not much has really changed. Team Green will leverage a wider memory bus (and external bandwidth) at the higher end, while Team Radeon will be limited to a 256-bit design (and use the Infinity Cache to make up for it). Furthermore, the next-gen RTX 40 series GPUs will retain Micron’s GDDR6X memory while AMD will stick to the more mainstream GDDR6 standard.
From what we know about Navi 31, it’ll be a chiplet design with two graphics compute dies (GCDs) and a massive L3 cache shared between the two in the form of a cache bridge. Each GCD will be roughly equivalent to the Navi 32/33 die with a bus width of 128-bit, resulting in an overall external bus width of 256-bit, same as the RX 6800 XT/6900 XT.
The Navi 32 and 33 GPUs powering the Radeon RX 7800 XT and 7700 XT will be paired with a 192-bit and 128-bit bus, respectively. These are likely going to be based on a monolithic design, both being faster than the present fastest RDNA 2 part. The core clocks should get a notable boost with RDNA 3, courtesy of TSMC’s 5nm (N5) node.
The same can be said for the RTX 40 series. They’re also speculated to adopt TSMC’s N5 process across the board, gaining process parity with their Radeon counterparts. Greymon expects a clock speed of over 2GHz (I’m guessing the base clock), with boosts approaching 3GHz. This isn’t one bit surprising as the Radeon RX 6000 parts based on TSMC’s 7nm already hit these targets. The N5 node should push it up a notch.