On November 15 (US time), Microsoft published online the Xbox 25th anniversary magazine, which it had first distributed at FanFest in Seattle. Microsoft had already announced at the Game Developers Conference (GDC) in March that the next-generation Xbox, Project Helix, will play PC games. In the magazine, Jason Ronald, who leads development of the next-generation console, explains in more detail how the interface will be designed so that players can start playing easily, PC games included. He also describes a plan to use machine learning to make rendering more efficient as component prices rise. Materials from Windows-based handhelds and development resources for the current Xbox show how Microsoft is making PC games easier to use, and how machine-learning-based rendering works in practice.
Cutting the effort needed to launch a PC game
In the magazine, Ronald explains that Helix is being designed as a home console that also supports PC games. Through a dedicated interface, players will be able to find and launch games intuitively and quickly, while still being able to change detailed performance settings when needed. He cited the ROG Xbox Ally, a handheld that shows PC games and console games on the same screen, as a precedent. Feedback from users of that device, he said, also influenced the design plans for Helix.
On the ROG Xbox Ally, which runs Windows 11, games installed from other PC stores also appear in the Xbox app's unified library and can be launched from there. In the ROG Xbox Ally update that Microsoft introduced in May, video output settings were also simplified when the device is connected to a compatible TV via a dock. When playing on a TV with a controller, this reduces the hassle of switching screens between stores to find a game, or of adjusting display settings every time you connect.
However, whether you can play a game you have already bought on another device at no extra cost depends on the terms of use for each title. With Microsoft's Xbox Play Anywhere, if you buy an eligible digital version from the Xbox or Windows store, you can download and play it on compatible devices at no additional cost. For example, you can download a supported title you started on your living-room Xbox to a Windows device and continue playing with the same account. Not only save data but also achievements carry over.
Efforts to make past Xbox games playable on PC are also moving forward. On November 12, Microsoft added 21 original Xbox games that can be played on Windows 11 in supported markets, expanding the lineup from four titles to 25. A download is required to play, and some titles require a purchase or a Game Pass subscription. All 25 titles support mouse and keyboard controls. Microsoft is not only making it easier to launch games from a PC, but also preparing each title's operating environment and input methods.
Which rendering processes does neural rendering make more efficient?
On page 17 of the magazine, Ronald points to neural rendering as a technology for improving graphics quality without relying solely on more computing power and memory, against the backdrop of rising component prices. It is a technique that builds a neural network into part of the process of generating images. At GDC in March, Microsoft had said that Helix will use a custom AMD SoC, co-designed with next-generation DirectX and FSR in mind. AMD's Jack Huynh announced "FSR Diamond", a technology optimized for Helix that will be built into the Game Development Kit (GDK) and use machine learning for super resolution, multi-frame generation, and ray reconstruction.
Upscaling is a technique that reconstructs an image a game rendered at a lower resolution into a higher-resolution one. According to AMD's technical documentation on FSR Upscaling for the current Xbox, it uses not only image color and depth but also motion vectors, which indicate the direction in which objects move on screen. By also referencing information from previous frames, it can start from fewer pixels than rendering every pixel at the output resolution from the beginning, while restoring detail according to motion. However, in areas that appear for the first time from behind an object because of camera movement, the information available from previous frames is limited. Microsoft's GDK sample also lays out the specifications for the input data required for upscaling.
Ray reconstruction is a technique that processes components found by calculating the paths of light rays, such as reflections and shading. When the number of rays traced in ray tracing is small, the results vary and noise appears in the image. AMD's ray reconstruction technology for current PCs applies a filter whose weights are determined by machine learning to the ray-tracing results, removing the noise. This process is a separate step from upscaling, which raises image resolution.
Frame generation, meanwhile, is a technique that creates new display images between the frames a game actually renders. As explained in AMD's FSR SDK development documentation, it is necessary to distinguish between how often the game renders frames and how often frames, including generated ones, are displayed on screen. Even if frame generation makes on-screen motion smoother, that alone does not make the game respond faster to input. To evaluate control performance, you need to measure the delay between a player's input being reflected in game processing and appearing on screen.
These machine-learning-based rendering techniques also require computation time to run the trained model and memory to hold intermediate data. For example, lowering the internal rendering resolution reduces rendering load, but a step to reconstruct a high-resolution image is then added. The GPU time actually saved depends on the difference between the rendering work that was skipped and the reconstruction work that was added, and it also varies with the image quality sought.
A 1.11 ms difference in the current Xbox developer sample
On November 4, Microsoft and AMD announced that FSR Upscaling 4.1.2, which uses machine learning, is now available for game development on the current Xbox Series X|S. Developer documentation published by Microsoft lists the GPU processing times when running an 8-bit integer (INT8) model in the GDK sample. Under the condition of 4K output on Xbox Series X, upscaling alone took 3.86 ms with the Full model and 2.75 ms with the Lite model. Comparing on the same console and output conditions, the Lite model is 1.11 ms faster than the Full model.
When displaying video at 60 fps, the time allotted to each frame is about 16.7 ms. The 1.11 ms difference produced by upscaling is not negligible when considering how much time the GPU has for each step. However, whether the time saved can be used for other rendering work such as lighting and reflections also depends on the order of processing on the GPU and the timing at which requests from the CPU arrive. These figures are measurements of the upscaler alone, and how much time can be secured for the game as a whole has to be measured across all processing.
Separately from processing time, the working memory used by the models can also be checked. The GDK sample's readme records that, with the "Quality" setting that upscales from an internal resolution of 2560×1440 pixels to 4K, the Full and Lite models used the same amount of memory. This tally does not include the memory for the swap chain, which displays rendered frames on screen in sequence. According to AMD's technical documentation, when input and output resolutions are the same, both models reserve working areas of the same size to hold intermediate images and the like. In other words, in this sample, switching the model to Lite shortened GPU processing time but did not reduce working memory usage.
When comparing image quality, video with the camera in motion matters as well as still images. Examples include whether thin fences and distant outlines stay stable in motion, and how detail in newly revealed areas from behind objects is reconstructed. By moving through the same game scene and checking whether the desired image quality is maintained, you can evaluate how much of the headroom created by shorter processing time can be redirected to other rendering work.
What the 2027 developer alpha hardware could reveal
Microsoft plans to begin shipping Helix developer alpha hardware in 2027. In the magazine, Ronald also mentions a vision of carrying over the game library and progress from four generations of Xbox. To judge whether players can actually move over from their current setups, the range of supported PC stores and older titles matters, as do the terms of use for each title. For example, even if a game appears in the unified library, whether a title bought from an external store can be launched as is, or whether you can continue from existing save data, is a separate question. The general release date and price have not been announced at this point.
For graphics performance, comparisons should use the same scene in the same game with a unified output resolution. When only the machine learning model is changed, the internal rendering resolution should be matched too, and when the internal resolution is changed to suit the rendering method, each setting should be stated. Image quality in moving video should then be compared, and the GPU time and working memory of each process measured. If frame generation is used, not only the displayed frame rate but also the delay until input is reflected on screen should be evaluated. If GPU processing time can be reduced while maintaining the desired image quality even in moving scenes, developers can redirect that headroom to other rendering work such as lighting and reflections.
