Today I am not going to read a specific paper. Rather, I watch a talk about some innovative ideas and oppotunities for computer architecture. The vedio of the talk can be found here
A Picture is Worth 1000 Bytes, Everything Else is AI
This talk propose Feature-first visual computing for mobile-cloud computing. On the mobile device, the app decompose visual contents into learned features from AI and send the “compressed” features to cloud. The cloud will do the operations on these small set of features and send the “compressed” results back to device. Then the device will “decompress” the results from cloud and generate final visualized results.
Advantages: (1) Less power uploading and dowloading full-size images/videos. (2) Good for privacy Drawbacks: (1) The AI-compressed features would make mistakes (2) ethics/ deepfack, etc (3) True data information will loss
Extreme Memoization: Everything in a LUT!
Although the Moore’s Law has come to an end, in contrast, memory and network are scaling up rapidly. This talk propose the idea of Extreme Memoization - store most computation performed and share it globally rather than recomputing. The memoizaton energy be will amortized.
Transform the function to an IR graph, and store the key-value pair in the global LUT. But the storage requirement for the global computation is tremendas, approxiamtion and DNA storage would help.
But I feel that the access to a really large LUT would consume a lot more energy/ latency compared with running locally.