AMD highlighted its GPU roadmap, as well as some further details on its future RDNA 3 architecture, at its financial analyst’s meeting. While the corporation didn’t go into great depth, what was revealed provided enough fodder for contemplation.
AMD claims that its RDNA 3 chips will be available before the end of the year, at which point at least some of the new cards will likely make their way onto our best graphics cards list and assist in reorganising the top of our GPU benchmarks hierarchy.
RDNA 3 is expected to boost performance per watt by more than 50%, according to AMD. As is customary, this must be viewed with caution, as performance per watt is a curve with a wide range of possibilities. Nvidia stated that its Ampere design gave 50 per cent more performance per watt than AMD’s rDNA 2 architecture, whereas AMD claimed that its RDNA 2 architecture delivered 50 per cent more performance per watt than its previous Turing architecture. While both assertions were true in some instances, numerous comparisons revealed far smaller benefits.
The RX 6900 XT is up to 112 per cent quicker and uses 44 per cent more power than the RX 5700 XT. This translates to a 48 per cent increase in performance per watt (perf/watt), which is near enough to 50% that we won’t quibble. Other comparisons, on the other hand, offer different outcomes. The RX 6700 XT is around 32% quicker than the RX 5700 XT while using roughly the same amount of power, or, for a truly horrible example, the RX 6500 XT is 22% slower than the RX 5500 XT 8GB while using 29% less power. That’s only a ten per cent increase in net perf/watt.
The RTX 3080 is around 35% quicker (at 4K) than the RTX 2080 Ti, but it consumes 28% more power. That barely equates to a 5% increase in raw perf/watt. The RTX 3070, on the other hand, is around 23% quicker and uses 11% less power than an RTX 2080 Super, resulting in an almost 40% increase in performance per watt. To give one last example, the RTX 3060 is 33% quicker than the RTX 2060 while consuming 7% more power, resulting in a 24 per cent gain in performance per watt.
The point is that AMD’s boasts of a 50% increase in performance per watt are only a best-case scenario. RDNA 3 might be 50% faster while using the same amount of power as RDNA 2, or it could be the same performance while using 33% less power. In general, genuine items will fall somewhere in between those two extremes, and it’s reasonable to say that not every comparison will show a 50% increase in performance per watt.
Other RDNA 3 details include the following, which we already know: It will make use of a 5nm process (almost certainly TSMC N5 or N5P). It will also include “enhanced multimedia capabilities,” such as support for AV1 encode/decode. DisplayPort 2.0 connectivity is included in AMD’s RDNA 3, which was previously rumoured.
We’ve heard a lot of rumours regarding the architecture, which consists of a redesigned compute unit (CU), AMD’s basic building piece for RDNA GPUs. The most likely scenario is that AMD will add additional computational pipelines, similar to what Nvidia did with Turing and Ampere.
Turing has two pipelines, one for FP32 and one for INT32, and Ampere introduced FP32 capability to the INT32 pipeline, thereby doubling the FP32 compute per streaming multiprocessor (SM) — Nvidia’s version of the CU. If that’s the case, RDNA 3 will almost certainly outperform RDNA 2 in terms of theoretical computational performance. It remains to be seen how this will affect actual performance.
AMD also offers a next-generation Infinity Cache, which has been shown to improve overall performance on the various RDNA 2 GPUs
This may be a huge cache, or it could be tweaked in some other way. We expect AMD will switch its top-tier RDNA 3 GPU to a broader memory interface, perhaps 384-bit, based on other predicted architectural improvements, but the firm has yet to remark on this.
Finally, AMD released the above GPU roadmap, which shows RDNA 4 as being scheduled for availability before the end of 2024. As you might expect, AMD didn’t reveal anything about RDNA 4 other than its inclusion on the roadmap and the “Navi 4x” naming scheme. It’ll very certainly use a 3nm process technique, or something similar under a different name. Based on prior launches, late 2024 is a safer prediction for the actual launch than early 2024.
AMD also restated its ambitions for an RDNA 3 launch in 2022, which is encouraging, albeit it might be a very limited launch that still qualifies technically
We’ll have to wait and watch whether AMD releases numerous graphics card levels before the end of the year. There’s still a lot we don’t know about RDNA 3, such as the number of cores, clock rates, and other architectural changes. AMD’s cDNA design features another GPU family, which we’ve examined separately, but there hasn’t been much overlap between RDNA and CDNA thus far.
Although the GPU cores are identical, RDNA has ray accelerators, and CDNA has matrix cores. RDNA 3 could follow in the footsteps of Intel Arc and Nvidia RTX by including a tensor/matrix core, but AMD has yet to make any such announcements.