3 ms·

Is inference fast enough to not have latency issues?

• wlesieutre 8y ago
  I assume there's some latency since it would need a fully rendered frame to operate on, but other postprocess AA techniques would suffer the same issue. Not a graphics expert, but I think only doing real supersampling would happen during the initial frame render instead of in series afterward, and that's going to tank performance much further than DLSS (in terms of individual frame time rather than latency, but much more significant overall). The DLSS neural network is trained off of renderings with 64x supersampling and tries to apply that to a low resolution image. In practice I wonder if there will be situations where the neural network screws up and makes strange artifacts, but the demos look impressive.