Energy implications of upscaling and frame generation in games

Sammanfattning

Gaming has a significant impact on residential electricity use, a cost that may increase with the introduction of new applications of machine learning, with neural networks known to have a large energy impact in other fields. This study performs an exhaustive analysis of energy usage of native rendering compared to modern and AI-based upscaling and frame generation technologies in games running on a mid-tier Nvidia GPU. It includes all single-player games (except one sequel in the same franchise) that as of April 15 offered a choice between the two frame-generation capable upscalers, Nvidia's own AI-based DLSS and AMD's FSR. Intel's XeSS was studied in any included titles that supported it. The study shows that native rendering efficiency varies widely across games and does not scale proportionally with resolution. Upscalers increased efficiency across the board, where FSR Ultra Performance with frame generation on average increased performance per Watt by 260 %, while DLSS reached the highest peak of 433 % gains in Cyberpunk 2077 with ray-tracing. When running uncapped, the efficiency gains matched with an increased framerate lead to a higher energy consumption of up to 6 %, except for DLSS frame generation that saved up to 7%. For users applying a framerate cap, substantial savings were observed across all upscalers with 19 – 49 %. The study additionally analyses upscaling cost by comparing the upscalers internal resolution to native rendering at the corresponding resolution; on average costing 0.6 Joules per frame. It further suggests methods for uncovering the cost of DLSS inference using CUDA tools and microbenchmarks. While the typical use case of upscalers, combined with the unknown training cost, leads to higher energy consumption, it's possible that their rumoured inclusion on future consoles can lead to both better performance and possible energy savings through commonly applied framerate caps.