All performance testing was done on a Radeon RX 9070 XT at 1440p resolution with maximum game settings. Frame rates were measured after the FSR 5 model had fully converged (10+ minutes of gameplay).
FPS Impact by Quality Mode
The performance overhead is consistent with AMD's claims: roughly 8–10% for Quality, 4–5% for Balanced, and 2–3% for Performance. These numbers are remarkably stable across different games, which makes sense—the AI core workload is largely independent of what the shader units are doing. Whether the game is GPU-bound or CPU-bound doesn't significantly affect FSR 5's overhead, because the neural rendering runs on dedicated hardware.
VRAM Usage
VRAM overhead matches AMD's specifications: approximately 400 MB for Quality, 250 MB for Balanced, and 150 MB for Performance. This is on top of the game's own VRAM usage. On the RX 9070 XT with its 16 GB of VRAM, this was never an issue, but it could become a concern on 12 GB GPUs in VRAM-heavy titles at 4K. The per-game pre-analysis cache stored on disk is typically 20–50 MB per game.
Level Load Times
The one-time texture, mesh, and shader pre-analysis adds a small amount of time to level loads. In our testing, the overhead ranged from 1.2 seconds (F1 25, relatively few unique assets) to 3.1 seconds (Cyberpunk 2077, massive open world with extensive asset streaming). On subsequent launches, the cached analysis data is reused and the overhead drops to near zero—typically under 0.3 seconds for cache validation.
This is a one-time cost per game that most players will barely notice, especially since modern games can already have substantial load times, even on NVMe SSDs.