3 ms·
Throwing ALUs and load/store units at the problem, specifically, seems like a great and obvious idea in retrospect. It seems to have become conventional wisdom.
by ahartmetz 4y ago
Throwing ALUs and load/store units at the problem, specifically, seems like a great and obvious idea in retrospect. It seems to have become conventional wisdom. They don't take much area compared to all the scheduling and bookkeeping hardware that tries to keep them busy.
- flatiron 4y agoAnd now we are paying the price of throwing scheduling hardware at the problem…
- ahartmetz 4y agoI was thinking specifically of Apple's M1 and recent upgrades from AMD and Intel, which all "go wider" (more execution units) and it works well for them. In some cases, these units can't even possibly be fed in steady state, only e.g. while catching up from some misprediction or cache miss.