At their 2022 Financial Analysts Day, AMD made some big promises for the GPU market. AMD has claimed that their new 5nm RDNA 3 GPU architecture will deliver a 50+% performance-per-watt uplift over today’s RDNA 2 GPUs. That means only one thing, that AMD’s next-generation GPUs are going to be VERY powerful.
Assuming that AMD will keep the power limits of their RX 7000 series RDNA 3 GPUs the same as their RDNA 2 models, that means that AMD’s next-generation flagship will offer users 50+% performance gains in most gaming workloads. If AMD increases their flagship GPU’s TDP, even larger performance gains should be expected. If this is true, RDNA 3 will be a game-changing architecture for AMD.
RDNA 3’s new features
With RDNA 3, AMD will be moving from TSMC’s 7nm lithography node to an unnamed 5nm lithography node. Beyond that, AMD will be using an “advanced chiplet package” to make their next-generation flagship more cost-effective than a monolithic GPU design. With RDNA 3, AMD’s bringing the cost benefits of chiplets to the GPU market. Chiplets is a core aspect of AMD’s success with Ryzen, so seeing Radeon move to chiplet-based GPU designs is a great thing to see.
With RDNA 3, AMD has promised users rearchitected compute units and an optimised graphics pipeline, signalling that we should expect IPC benefits from AMD’s newly redesigned GPU compute units. This is likely a core area where AMD has enhanced RDNA 3’s performance/watt characteristics. Beyond that, these re-architected compute units could, potentially, deliver dramatic increases to AMD’s ray tracing performance with RDNA 3.
RDNA 3 will also double-down on AMD’s Infinity Cache technology, a technology that AMD pioneered with their RDNA 2 graphics cards. What this change means remains to be seen, but the good news for gamers is that Infinity Cache is effectively an anti-crypto-mining technology, as Infinity Cache allows AMD to achieve higher performance levels with a smaller memory bus, and most cryptocurrency mining operations are very bandwidth-limited and are too large for GPU caches.