Associate
- Joined
- 29 Jun 2016
- Posts
- 529
So if a game level has listed 10gb of textures possible, but each frame concurrently uses only 2gb of textures they propose to juggle from the slower storage the 10gb into the 4gb of memory so that it is no disadvantage. Its brains not speed so much, deep learning related ? I dont know how effective but nice if it works.
Deep learning is so hot right now. Gotta get those buzz words in.
It can give you predictive non-linear responses to stimuli, but I fail to see how it can save you memory space. I mean you could load a network and weights trained for each game/workload, but it just seems mental.
My gut feeling is that their deep learning "it is almost human" marketing garbage is due to the strong FP16 performance @ 25TFlops, which is ideal for use with big flippin' neural networks.
Last edited: