AMD outlines RDNA 3 GPU roadmap featuring 5 nanometer chiplets launching in 2022.

AMD announced its RDNA 3 architecture during a financial analysts briefing, revealing that the chips will debut before the end of the year and will be projected to provide more than a 50% uplift in performance per watt. The company noted that AMD's RDNA 2 architecture was also supposedly a 50% improvement in performance per watt, as was Nvidia's claim that its Ampere architecture delivered 50% more performance per watt than its previous Turing architecture. However, actual performance comparisons tell a more nuanced story. Based on performance and power testing, the RX 6900 XT is up to 112% faster than the RX 5700 XT while using 44% more power, representing a 48% increase in performance per watt. Other comparisons yield different results: the RX 6700 XT is about 32% faster than the RX 5700 XT while using basically the same amount of power, while the RX 6500 XT is 22% slower than the RX 5500 XT 8GB while using 29% less power, yielding just a 10% net performance per watt improvement. Nvidia's cards show similar variation: the RTX 3080 is about 35% faster than an RTX 2080 Ti at 4K but uses 28% more power for a 5% raw performance per watt improvement, while the RTX 3070 is about 23% faster than an RTX 2080 Super while using 11% less power, nearly a 40% improvement in performance per watt. AMD's claims of 50% higher performance per watt represent a best-case scenario that could mean RDNA 3 is 50% faster while using the same power as RDNA 2, or the same performance while using 33% less power.

AMD confirmed several technical specifications for RDNA 3, including 5nm process technology, almost certainly TSMC N5 or N5P, and support for "advanced multimedia capabilities," including AV1 encode/decode support. DisplayPort 2.0 connectivity is also included, consistent with earlier rumors. The architecture features a reworked compute unit, the main building block for AMD's RDNA GPUs, with the most likely scenario being something similar to what Nvidia did with Turing and Ampere, adding additional compute pipelines. If correct, RDNA 3 will likely have a big jump in theoretical computational performance relative to RDNA 2. AMD also promises a next-generation Infinity Cache, which has proven effective in increasing overall performance on various RDNA 2 GPUs, and the company is expected to move its top-tier RDNA 3 GPU to a wider memory interface, probably 384-bit, though it has not commented on this officially.

Perhaps the most interesting aspect of RDNA 3 is AMD's use of advanced packaging technologies combined with a chiplet architecture, similar in some ways to the approach seen with the X670/X670E chipsets and AMD's proven CPU chiplet architecture. One likely configuration would have a base GPU chiplet targeting the midrange performance segment that can be linked together with one to three additional chiplets via a high-speed interconnect, functioning somewhat like SLI and CrossFire but without alternate frame rendering. The operating system would only see one GPU despite multiple chiplets, with the architecture managing data sharing between them. AMD has already demonstrated this capability with its Aldebaran MI250X data center GPUs, and Intel appears to be taking a similar approach with its Xe-HPC designs. AMD also revealed details on its data center CDNA 3 GPUs that suggest those will have up to four GPU chiplets, though these are compute-focused architectures rather than designs focused on real-time graphics performance.