A new leak regarding AMD’s unreleased RDNA3 GPU architecture has emerged. More to the point, the alleged details pertain to its Navi 31 and 32 GPUs and their internal specifications.
According to a report by Angstronomics, the site reportedly got its hands on the final specifications for each of the RDNA3 Navi 3X GPUs, from 31 to 33. From the looks of it, it appears that AMD made a 3D stacked variant with double the capacity. Sadly, it found the performance benefits to be underwhelming and it is due to that reason that it supposedly made the top-tier Navi 31 SKU, along with a watered down version with just 80MB of the proprietary cache.
There’s also new information regarding Navi 32 as well. Apparently, the GPU is expected to have 30 WorkGroup Processor, instead of the initial 32 that was rumoured. In other words, the GPU itself may offer as many as 7680 Stream Processors, and not 8192 that was initially reported.
There is also word about a new reference cooler design for RDNA3 its corresponding Radeon cards and more specifically, the Navi 31 cards. It supposedly still has three fans but the height of all cards are expected to be a little taller than the current Radeon 6000 series card. Additionally, there’s also a “3-stripe accent” near the cards’ dual 8-pin PCIe power ports. While there are no images or artist’s impressions to support the last claim, the use of power delivery suggests that AMD isn’t going with the PCIe Gen5 power delivery method like its direct rival, NVIDIA, may or may not adopt with the launch of its next-gen RTX 40 series GPUs.
AMD confirmed earlier that RDNA3 was well underway and that it would be based on the 5nm die lithography, derived from TSMC’s process node, naturally. In addition, the new GPU architecture would feature a new advanced chiplet packaging and a “Rearchitected Compute Unit”.
In that same announcement, AMD also said that its RDNA3 architecture would ship out with a next-gen Infinity Cache. So, in that aspect, this rumour does, to some measurable degree, hold water.