3 ms·
Neural network for memory access huh? Interesting!
by deepnotderp 10y ago
Neural network for memory access huh? Interesting!
- agumonkey 10y agoI wonder how this will affect the software side. Programmers and compilers.. how do you interface with a NN based prefetcher to avoid regression and hiccups.
- snerbles 10y agoIf it's a truly adaptive branch predictor, one could naively believe it will yield good results for most inputs. Even if it's an absurdly deep pipeline with a cache miss penalty worse than the Pentium 4, it would be premature optimization for the vast majority of programs.
- filereaper 10y agoMost processors only let compilers give a hint of strongly taken or strongly not-taken. I don't have any numbers which show regressions, but a NN should be able to take the compilers hint.
- PeCaN 10y agoI believe Intel has been doing this for a while, or at least it's definitely not the first time I've heard of the idea. ‘Neutral network’ is more of a marketing term, though somewhat accurate in that it's a number of perceptrons doing statistical branch prediction. No multilayered networks (as far as I know) have been used in any CPU to date.
- Tuna-Fish 10y agoNo. Perceptron predictors are very much an AMD thing, and they have been doing them for a while. Intel uses more traditional predictors but with larger buffers. Both approaches have situations that are more or less favorable to them.