In other words, 3.25x the effective bandwidth of a 256-bit cache. One that can turn 512GB/s bandwidth into 1,664GB/s. The result is a cache subsystem that, AMD says, offers an effective bandwidth far exceeding the physical 256-bit one. "An architecture that's easy for the industry to adopt" Laura Smith, senior director, RTG So that's why AMD took its CPU cache and made it increasingly more parallel in the Infinity Cache. That's great for a CPU, but when it comes to the highly parallelised workloads of a GPU that simply won't cut it. In fact, it's the work that AMD's engineers carried out on its EPYC server chips that led to such a cache technology making its way to RDNA 2 in the first place.Ĭache is effectively organised into serialised banks within EPYC, one after the other. It's a close relation to Infinity Fabric, the interconnect used throughout Zen 2 and Zen 3 CPUs. Infinity Cache is what you might call a cure-all for AMD's GPU issues of the past- its gaming panacea, so to speak. Instead, AMD is incorporating a new concept with RDNA 2: Infinity Cache. That's also why you'll not see HBM2 memory anywhere near RDNA 2's hallowed PCB either. Two things AMD of all companies knows it can't get away with. From here AMD's building out with an eye on maintaining a level of efficiency throughout, and not blowing the entire power budget on a 512-bit bus, I'm told.ĪMD really wants people to know bandwidth is not all about bus width-forget what ya heard.Ī 512-bit bus, AMD explains, is costly and large.
From there, it's all about scaling up an architecture that just doesn't waste any juice.Īt a base level you're already getting more performance out of the same components with RDNA 2, but that doesn't make for a lax approach on a uniquely architectural level. It's all in order to ensure that a maximum benefit and power saving is being achieved at a transistor and clock level. These optimisations go down to the very capacitance of transistors as they switch. To further help the architecture along AMD has built in further capacitance in alternating current (CAC) optimisations right down to the metal with RDNA 2. The RX 6900 XT doubles the CUs, all of which run up to 18 percent faster, for just 33.33 percent greater power. This first wave card delivers 40 CUs at 225W with a top clock of 1,905MHz. For a clear cut example of the initial power savings and 7nm optimisations, just look to the first-generation RDNA top dog: the RX 5700 XT. That cross-pollination of ideas and methodologies between Zen and RDNA 2 starts here: In order to improve performance while retaining roughly the same process node, AMD has incorporated custom libraries, tools, methodologies, and structures used to build Ryzen 5000-series CPUs into RDNA 2 in order to run it at higher frequencies or the same frequencies for less power. RDNA 2 pulls no punches in either, offering 30 percent higher frequency for the same power draw as the previous generation, or 50 percent greater power savings while holding the same frequency. With a continuation of the 7nm process, you'd be easily fooled into thinking that there's little to talk about when it comes to performance per watt and clock speed. (Image credit: AMD) AMD RDNA 2 architecture