Microsoft puts numbers on Xbox’s new AI upscaling
Illustrative image: Martin Martz on Unsplash
Microsoft’s 4 November developer update puts a GPU cost on AMD’s machine learning upscaling for Xbox Series X and Series S, with a sample offering Full and Lite INT8 models. [1]
On Series X, the sample takes 2.75 ms to produce a 4K output with Lite and 3.86 ms with Full. A game targeting 60 frames per second has 16.67 ms between frames, calculated as 1,000 divided by 60, so those upscaling times are equivalent to roughly 16.5% and 23% of that interval. These are timings for the sample’s GPU work, rather than measurements of a game’s final frame rate. [1]
FSR reconstructs the output from a lower resolution image, using information supplied by the game engine. AMD’s integration documentation calls for colour and depth buffers, motion vectors describing movement across the screen and small offsets in the camera’s sampling position. In its Quality mode, a 4K output starts with a 2560 × 1440 image, about 44% of the output pixel count. Rendering fewer source pixels can reduce graphics work, with reconstruction adding a separate processing stage afterwards. [2] AMD’s Performance mode instead starts at 1920 × 1080 for 4K output, a quarter of the output pixel count. That reduces the source image further, leaving the reconstruction algorithm with less information from which to recover detail. [2]
Thin objects and particles are particularly revealing when the camera moves, because lost detail, flicker and ghost trails can be much harder to disguise than softness in a stationary scene. AMD says its newer model improves these areas compared with FSR 3.1. Those claims describe what the algorithm is designed to improve; choosing between Full and Lite still calls for examination of moving scenes alongside the timing measurements. [2]
AMD’s wider SDK uses a smaller INT8 model for selected processors with more constrained integrated graphics, extending machine learning upscaling to laptops and handhelds. The company says this lightweight model is less than half the size of its original INT8 model, while more capable GPUs use the full model. This is inference running within the existing graphics hardware, with the model chosen to suit its resources. [3]
At 1080p output on Series S, Lite takes 1.9 ms and Full takes 2.9 ms in Microsoft’s sample. Developers can compare both with FSR 3 before integrating the selected option into their games. [1]
