The Xbox Series X's Velocity Architecture features many revolutionary new technologies and architectural optimizations, one of which is Sampler Feedback Streaming, which can dramatically reduce texture pack file sizes by eliminating redundancy.
The Xbox Series X is a game-changer. The sum of its parts--whether it be the higher-end Navi GPU, Zen 2 CPU, and powerful PCIe 4.0 SSD hardware or the new software/OS/architecture optimizations--are all massively synergized to work together. Sampler Feedback Streaming is one of these parts, and it's featured in the new Velocity Architecture, a banner that's focused on making the Xbox Series X fast and efficient.
We've already outlined the basics of Sampler Feedback Streaming in our in-depth Xbox Series X SSD video, but now we have a developer's perspective on the topic. Despite Microsoft's transparency on the console, devs have been markedly silent on how the Xbox Series X's new features will benefit them. But now Team Blur Games' lead designer Gavin Stevens offers some golden insight.
Stevens recently posted a huge Twitter thread that outlines just how remarkable Sampler Feedback Streaming is, and what it means for developers.
In short, SFS gives devs much more granular and fine-tuned control over how data is prepared and delivered to the Navi GPU for rendering. SFS takes data that's uncompressed by the I/O system--in this case, the new DirectX 12 StorageAPI--and then helps streamline the data asset pipeline as the content is delivered on through to the GDDR6 RAM, which is then passed to the CPU for processing and the GPU for rendering.
This pipeline is called data streaming, not to be confused with cloud/game streaming, because the data is streamed from storage to the memory pool and SoC.
So how does this tangible affect games development?
SFS can eliminate redundant mips and significantly reduce overall texture data sets, which is extremely important for 4K (or even 8K) games. As Stevens explains, textures require mipmaps, or downscaled versions of the textures. If textures don't have mipmapping then the scene will look too stark and unrealistic. Mipmaps blend the textures together to make them look more natural and easier on the eyes.
Essentially SFS will only load the required mip at any given time for any given solution. The result is a lot less data that's loaded into the RAM, which means less congestion and more space in the memory pool and less taxation on the I/O storage throughput that's constantly moving data.
Here's how Stevens explained it (be sure to read the whole thing as it's pretty illuminating):