| Topic List | Page List: 1 |
|---|---|
| Topic | Computing power may increase 100x in the near future (shitty source) |
| Sahuagin 07/22/20 6:17:32 PM #6: | Sahuagin posted... the other paper is discussing a method of implementing photonic TPUs.after reading a chunk of the paper, it is specifically a method of implementing parts of TPUs to allow fast matrix multiplication (using existing optical methods). apparently matrix multiplication is an important part of neural-network machine learning, but has a computational time complexity of O(n^3) (the time it takes to compute increases with the cube of the width of the square matrix). (there is also a problem I keep reading about recently regarding von Neumann architecture and the CPU-memory bottleneck, which is not present in GPU and TPU architectures.) using electronic computers and some kind of optimization technique you can get matrix multiplication complexity down to O(n^2.373), but the paper is describing in great detail an optical hardware implementation that can compute matrix multiplications (with one varying input matrix, and one relatively non-varying matrix) in O(1) time complexity (specifically less than 20 picoseconds). probably not worth going into much more detail than that. they use a germanium/antimony/selenium glass for the 'wires', with tungsten metal walls, and they have to use "heterogeneously integrated optimized photonic memory" for reasons, and you can see diagrams of the multiplexers and so on... I don't know what there is to doubt about this. it's great to see advances in both realms of hardware neural-network processors and optical processors. as for the thread title "Computing power may increase 100x in the near future", that's kind of misleading. the paper says they can increase TPU power efficiency by hundreds or thousands of times. this is not something that will directly apply to your home computer, but eventually this kind of work should lead to better home computers. --- ... Copied to Clipboard! |
| Topic List | Page List: 1 |
