NVIDIA may have hyped the crap out of its new Ampere GPU lineup, but you weren't expecting AMD to sit around and do nothing, were you? That's not how Dr. Lisa Hsu rolls. Team red has fired back with its RDNA2 Architecture for PC, set to come out in mid-November. So, let's dive right in with what to expect.
First off, let's talk about performance. AMD is initially releasing three cards, each of which is expected to compete with an equivalent RTX 3000 series model. At the top end is the $1,000 RX 6900 XT. And it's supposed to deliver RTX 3090-level performance but using significantly less power. And seeing as Ampere cards are notoriously power hungry, it could be an interesting option for enthusiasts that don't want to spend even more money upgrading their power supplies.
Next up is the RX 6800 XT, not to be confused with the 2005 vintage NVIDIA card of the same name. That's right, we're looping back on GPU names now. This card is going to go for 649, a little less than its direct competitor, the RTX 3080. One step down from that is the RX 6800 non-XT, which AMD suggests will be slightly better than the 499 RTX 3070, for 80 bucks more.
Unsurprisingly, AMD has put out a few benchmarks, which you can see on this chart impressive numbers to be sure but how did they do it? And how much skepticism should we view these first party numbers with?
RDNA2 is actually built on the same seven nanometer process as the previous generation. So performance improvements are primarily driven by changes to the architecture itself, instead of just cramming more transistors on a substrate. But with higher and higher resolutions becoming more and more mainstream, VRAM is becoming increasingly important. And AMD thinks a couple of innovations will give Radeon an edge in 4K.
The first is what they're calling, Infinity Cache, which is a small amount of memory directly on the GPU die itself. Although it only consists of 128 megabytes, it's very fast. AMD claims that the result is an effective VRM bandwidth of nearly 1700 gigabytes per second, a huge number compared to the GDDR6X standard used by the upper tier RTX 3000 chips.
However, the nominal bandwidth without Infinity Cache is the far more pedestrian 512 gigabytes per second. So it remains to be seen exactly how the high-speed cache will play out from title to title. But, AMD does claim it's responsible for nearly half of the performance increase from the previous generation.
The other big innovation is called Smart Access Memory. Which gives more direct VRAM access to the CPU instead of allowing the CPU to access the VRAM only in smaller chunks. AMD claims you'll get up to a 10% performance boost as a result, but the big caveat here is that you'll need both a Ryzen 5000 CPU and a motherboard with an AMD 500 series chipset in order to use it. So you'll miss out if you're using an older AMD platform or an Intel based system
Keep in mind that the benchmarks we've already seen had Smart Access Memory enabled. So the numbers may be a little less impressive, once independent testing on non-Ryzen 5000 systems is conducted.
Another major first for AMD here is Hardware Raytracing Support, through what it calls Ray Accelerators, which actually isn't a weapon that you'd see on the Starship Enterprise. They're actually similar to the Raytracing cores NVIDA introduced with the RTX 2000 series. However, AMD hopes that built-in de-noising technology can help it's Raytracing solutions look actually better than NVIDIA's.
There's also Direct X 12 Ultimate support, as well as compatibility with direct storage to reduce CPU overhead and a DLSS like feature called Super Resolution. Basically it all looks like a very worthy competitor to Ampere, with a few of AMD's own unique twists.