AMD RDNA5 Architecture Will Be A βClean Sheetβ Design, Leaker Alleges
AMD RDNA5 Architecture Will Be A βClean Sheetβ Design, Leaker Alleges
While weβre still months away from AMD potentially launching RDNA4-based GPUs, leaks surrounding the RDNA5 architecture have now surfaced. This leak didnβt mention any hard numbers, but there are a few important points worth discussing.
The leak originated from leaker βwjm47196β on Chiphell forums, who hasnβt been active for a while and has decided to provide more look from the inside on RDNA5, but also decisions surrounding RDNA4βs absence from high-end GPU segments. Hereβs the full translation:
AMD never planned to give up on high-end (GPUs), due to various reasons RDNA3 failed to achieve the performance expectations, causing the frequency/power curve to go out of control. This includes, but not limited to, conducting verification for the MI series (AMD Instinct) chip interconnects and sharing R&D costs. Originally there was a design with 192MB Infinity Cache which got cancelled due to cost and power consumption reasons.
RDNA5 will be a clean sheet design similar to how Zen works (could be a name change? this is uncertain), hence a stopgap was created via RDNA4 to thoroughly conduct case study and R&D.
RDNA4 is merely aimed to fix the βbugsβ from RDNA3, hence the performance is still most likely going toe-to-toe against the Radeon RX 7900 XT in a best-case scenario, though ray tracing has been improved.
Hereβs something to add, regarding AI software currently it looks like Xilinx is getting involved, building up the ROCm ecosystem is currently AMDβs top priority right now
We have learned from earlier rumors that AMD will not be competing in the high-end segment this coming generation, meaning NVIDIA will get full control of the flagship GPU segments for a good while. If this leak is anything to go by, the hiatus will be a short one β Team Red will be looking to return to the high-end provided their R&D efforts result in a highly competitive architecture that can properly compete with Team Greenβs GPUs.
AMDβs recent acquisition of Xilinx, an FPGA chipmaker, may pay dividends for its future chip designs as well. Aside from being mainly responsible for NPUs found in Ryzen AI, there is a solid possibility that the chipmaker will be implementing AI-related features into its future GPUs with dedicated accelerators and such. Currently, Xilinxβs focus on the ROCm software stack will allow AMD to challenge NVIDIAβs longstanding CUDA dominance.
Source: Videocardz
Pokdepinion: Time to keep expectations low for RDNA4 then.Β

