There is no disputing that AMD has made some giant strides in the enthusiast market, especially with its Zen CPU architecture, though also with its Vega GPUs. The latest incarnation of the latter is the recently announced Radeon VII with a 7nm Vega GPU inside. It's a beastly card in its own right, though NVIDIA still has the edge in overall performance with its own Turing GPU architecture. Will AMD ever close the performance gap? That is an interesting question to mull, in light of recent patent applications that AMD filed.
First let's talk about the performance of the Radeon VII. The only numbers we have to evaluate at the moment are AMD's own benchmark figures. If those are a proper representation of the card, then performance will be about the same as NVIDIA's GeForce RTX 2080 in rasterized games, and a bit faster in titles that use the Vulkan API.
That means we can reasonably expect the GeForce RTX 2080 Ti will still be the top dog in graphics, and whatever gap there is between it and the Radeon VII will only widen when enabling real-time ray tracing, a feature AMD's GPUs are not equipped to handle.
Now here's where things get interesting. A patent application titled "Stream processor with high bandwidth and low power vector registry file" has come to light, and it follows a previous patent published in May titled "Super single instruction multiple data (Super-SIMD) for graphics processing unit (GPU) computing."
Block diagram of one embodiment of a stream processor
Taken together, the patents outline how AMD might be thinking about its future GPU design, and how things could look beyond Navi.
I won't pretend to fully grasp all of the technical details, but the gist of it is that AMD is exploring a design that would result in less shared resources than its current Graphics Core Next (GCN) design. The new high bandwidth stream processors look to have more logic crammed inside, each with its own set of cache, buffers, and instruction queues.
Utilizing this design, AMD still might not be able to cram more stream processors or cores into its GPU architectures than it currently can, but it would open things up and potentially avoid performance bottlenecks that are inherent when sharing resources.
Hit the link in the Via field below to read an argument on how the patents might resemble NVIDIA's approach to GPU design. I'm not sure if that's actually true (it's refuted in the comments section of that article), but it's interesting nonetheless.
What I can say is that AMD's GCN architecture has been the company's driving force in GPU design for a long time now, having debuted in 2011. There have been rumors that Navi will be the last GCN-based architecture. These patent applications add some credibility to the rumors, though obviously we will have to wait and see what transpires.