3 ms·
So the basic idea is to do some calculations right inside the 4-18Kb memory block instead of going all the way to a CPU and back? First of all, why does this s
by kramerger 3y ago
So the basic idea is to do some calculations right inside the 4-18Kb memory block instead of going all the way to a CPU and back?
First of all, why does this sound familiar? Did early memories do something like this?
Second, how hard is it to map arbitrary complex NN to this architecture?
- PaulHoule 3y agoThere was STARAN would could accelerate operations such as predicting if planes were on collision courses in air traffic control https://www.cs.kent.edu/~parallel/papers/p405-batcher.pdf https://www.cs.kent.edu/~parallel/papers/p405-batcher.pdf
- touisteur 3y agoFloating point might be expensive on die room, right now I think it's mostly integer, bit manipulation, and compare-and-swap, which is already a lot and may help with quantized network. Curious whether someone implemented FP32/16 or some non-IEEE754 floats in there. My worry with all these archs is the ease of debug, tracing and grokking whatever the hell is happening. Disaggregation of compute will make programs very hard to write and understand...
- KenArrari 3y agoI think the main thing about PIM is that it typically doesn't make sense to do compute on the same die as the memory because it would be too slow to be useful (you're better off going to the CPU for that). But it can make sense when it's a very simple and repetitive calculation. ie: copy data from one location to another, bulk bitwise ops, etc.