Accidental Tech Podcast621: I'll Meet You Any Time You Want
https://atp.fm/621
Breaking news and in-depth coverage from our reporter at CES.
Stephen Tierney@atpfm RE DLSS speed.
DLSS used to use a convolutional neural network, which you can run at high frame rates on consumer HW. Here's Yann LeCun demoing CNNs in 1989 in near real time https://www.youtube.com/watch?v=FwFduRA_L6Q
NVIDIA makes further optimisations to squeeze their DLSS models so that they are sufficiently fast. They use techniques such as pruning, distillation and quantisation.
Commercial LLMs like ChatGPT don't undergo the same level of aggressive parameter optimisation.
