Microsoft has revealed some new data on how the Xbox Series X will utilize its advanced high-speed storage capacity and some of the GPU efficiency improvements we can expect to see. Some of these technological developments could have a direct bearing on the future of PC GPU technology, particularly if they were developed in association with AMD.
Microsoft’s latest blog post details what it calls the Velocity architecture. The first section of the blog only shows the data we already know, in terms of speeds and feeds: 12TFLOPs, 16GB of RAM,> 4x the CPU performance of the Xbox One, and so on. The Velocity architecture is comprised of four components: a custom NVMe SSD, hardware accelerated decompression blocks, a new DirectStorage API layer, and a new technology called Sampler Feedback Streaming. It is the latter that seems as interesting as a potential PC technology.
The SSD offers 2.4 GB / s of raw storage / 4.8 GB / s of compressed data, which is 50 percent of the rate that Sony has specified for the PlayStation 5. Microsoft can’t resist a bit of FUD, writing: “Traditional SSDs used on PCs often reduce performance as thermal increases or when servicing the drive.”
The first means that it has a terrible cooling system, while the second is not known to be a problem in most conventional storage workloads.
Next are the hardware-accelerated decompression blocks. This is where the new BCPack algorithm comes in, but the XSX also supports the industry standard LZ decompressor and can run both simultaneously. This is where the derived bandwidth of 4.8 GB / s from Microsoft comes from. This is a 100x improvement over current consoles and, according to Microsoft, it would require more than 4 Zen 2 cores to deliver in pure software.
The third piece of the puzzle is the new DirectStorage API. This is how Microsoft describes it: “We added a new DirectStorage API to the DirectX family, providing developers with fine-grained control of their I / O operations, allowing them to establish multiple I / O queues, prioritizing and minimizing latency of is”. Developers will have tremendous flexibility to take advantage of underlying hardware I / O performance.
These gains are important to the overall design of the XSX system. Unlike the previous consoles, the RAM load on the Xbox Series X only increases 2 times compared to the Xbox One and only 1.33x compared to the Xbox One X. That is much less than in previous generations; the Xbox 360 had 512 MB of RAM, while the Xbox One had 8 GB. That’s a 15,625x magnification from 360 to One, and a 2x magnification from Xbone to XSX, the best case scenario.
The only way to take another leap up in graphics quality without adding large amounts of system memory is to use SSD storage very, very efficiently.
Sample feedback transmission
Today’s GPUs use a technique known as mipmapping to optimize game textures for different levels of detail and resolutions. A player who is very far from an object will see a version of it, while approaching it will instruct the game to change to a different MIP map. Texture resolution has to be increased to provide higher quality images in a closer range, but the way MIP maps are designed requires a GPU to load a full MIP map into memory, even if it only needs a relatively small chunk little.
Determining how much RAM games actually use is a difficult task, even today, but Microsoft was able to use specialized hardware within the Xbox One X to measure how the GPU used texture memory internally. What they found was surprising: GPUs commonly access less than 1/3 of the texture data they require to load into RAM. This has some significant implications for how more efficient memory usage patterns could allow GPUs to consume less power or spend more than they consume to improve performance. Microsoft writes, “A single scene often includes thousands of different textures that result in significant loss of effective memory and I / O bandwidth utilization due to inefficient use.”
The XSX sidesteps this problem with a new technology that allows it to load sub portions of MIP maps instead of the entire texture. MS writes that this allows 2.5 times the effective I / O performance and memory usage, beyond the capabilities of the raw hardware. This technology could help explain how AMD is claiming a 1.5x increase in performance / watt with RDNA2. I’m not assuming there is an automatic link between the two, because Microsoft could have an agreement with AMD that reserves certain technology specifically for Xbox. But this could be part of the secret sauce that helps RDNA2 match or exceed Ampere’s performance.
Microsoft released the video above as part of the Velocity Architecture announcement, so check it out for more details. Now I’m curious to see what features the upcoming RDNA2 GPUs will and will not support, relative to what we see the Xbox X Series offering when it launches later this year.
Now read:
