Nvidia Will Mimic AMD’s Smart Access Memory on Ampere: Report
We haven’t even gotten to talk about the results of features like AMD’s Smart Access Memory, and Team Red’s competitors are already lining up to pledge support for an equivalent feature. SAM is AMD’s first-ever method of taking advantage of the fact that it owns its entire IP stack — CPU, GPU, and platform. According to AMD, it can use this fact to boost system performance when you combine an RDNA2 card and a Ryzen CPU with an X570 motherboard.
Specifically, AMD claims that it can give the CPU full access to GPU RAM rather than limiting the system to a 256MB aperture window for data transfers. We don’t know if the large Infinity Cache integrated into RDNA2 plays a part in this. No one has stuck a 128MB cache in a GPU before, so the idea that it could play a role in boosting data transfers in and out of VRAM isn’t crazy, especially since cache latency is presumably quite a bit lower than the time it takes to go out to GDDR6.
According to GamersNexus, Nvidia is capable of doing something similar:
Hard to fit in a tweet, but basically, they’re working on enabling the same feature as AMD Smart Access Memory (AMD GPU+CPU=Perf uplift) on both Intel and AMD. No ETA yet. Doesn’t look like it’ll be ready before RX 6000 launch, but we’ll keep an eye on development.
— GamersNexus (@GamersNexus) November 12, 2020
Nvidia claims that the feature is part of the PCIe specification, that it has the capability working internally, and that it is already seeing similar performance results. There’s no ETA on when the feature might be available in-driver.
AMD has not publicly disclosed whether Smart Access Memory depends on features beyond the ability to adjust the size of this aperture (Intel refers to this as the IGD Aperture Size), or if it is additionally enhanced by capabilities of Zen 3 or the B550 / X570 chipset. It would definitely change the framing of the feature if Nvidia were capable of activating it on both Intel and AMD systems — but if it is, the likely end result would be AMD activating the capability for Intel systems as well.
Historically, in order for capabilities to become common, both GPU manufacturers have to agree to use a common standard. Nvidia introduced ray tracing using Microsoft’s DXR, for example, but it was only when AMD added the capability to next-gen consoles and its own GPUs that it really began taking off in the mass market. If Nvidia and Intel can both take advantage of a feature that AMD currently only enables for AMD customers, the company will likely enable it on Intel platforms as well, or risk losing match-ups to its biggest rival.
AMD is unlikely to want to risk that outcome. It’s one thing for the company to enable additional performance by leveraging commonalities within its own ecosystem and another to artificially limit performance on competitor platforms.
No matter what, though, the end result seems to be a consumer win. If Nvidia successfully adds this feature, game performance goes up for Nvidia customers. If AMD responds by enabling it for Intel as well, performance goes up for AMD + Intel customers. If, on the other hand, Nvidia (or hypothetically, Intel) can’t match AMD’s gains or fundamental capabilities, we’ll have evidence that the company really is taking advantage of its cross-product IP in new and interesting ways.