The Twitter rumor mill is now calling 350mm² or slightly more for the Navi 31 GPU that will be part of the Radeon RX 7000.
Navi 31’s GCD (Graphics Complex Die) is said to be 350 square millimeters (or slightly larger). At least that’s what Twitter user Greymon55 wants to know, who always seems to know a lot about new graphics cards. 350 square millimeters (plus) is relatively small if you look at the general development in this segment, but it is also primarily the result of AMD’s efforts with an MCM design. The whole package will consist of GCD and MCD; the sizes of the MCDs (Memory Complex Die) and the package are not yet known. A rough comparison to the Navi 21’s monolithic design: It’s 520 square millimeters in size, but it’s also a monolithic design with caches and memory controller.
Radeon RX 7000: AMD confirms higher power consumption with RDNA 3
Meanwhile, the increasing amounts of shaders and larger caches would suggest that Navi 31 will increase overall. According to rumors, the entire package for the Radeon RX 7900 consists of a GCD and six MCDs. It is estimated that an MCD will be around 40 square millimeters, bringing the complete design to around 590 square millimeters. Nvidia’s AD102 is said to measure around 600 square millimeters as a monolithic design.
Ever larger caches consume more and more chip area, but from time to time savings can be made. With Navi 31, AMD is said to have thrown unnecessary ballast overboard, among other things. XGMI, Global Data Share (GDS), Legacy Geometry Pipeline or Legacy Scan Converter is what Twitter users call Kepler_L2. Of course, the refined manufacturing process in 5 nm also helps to counteract the growth in shaders and caches. While RDNA 3 has so far been traded as Navi 31 with 12,288 shaders, the chip(s) with 16,384 should still be an issue. AMD had already agreed to increasing power consumption because the competition also handles it that way. In the end, however, the fact remains that more computing power is currently not possible without higher power consumption, despite finer manufacturing processes and improved architectures.
Source: Twitter (Greymon55)