3 ms·
To a degree, that makes sense. Due to the end of dennard scaling, most high end raw compute makes more sense as more but simpler cores, and has for a long time
by monocasa 2y ago
To a degree, that makes sense. Due to the end of dennard scaling, most high end raw compute makes more sense as more but simpler cores, and has for a long time. For instance the blue gene supercomputers made of tons of pretty individually anemic PowerPC 4xx cores.
For battery powered devices, race to sleep is the current meta, where once you have some bit of heavy compute work, you power up a big core, run it fast to get through all of the work as quickly as possible, and get back to low power as quickly as possible.
- hajile 2y agoBecause clockspeed ramps power exponentially, there’s a limit to how high you can clock before the cost of racing outweighs the savings of running a short time. I believe I’ve read that Apple’s chips run under 3GHz unless their job runs longer than 100-150ms. I suspect that’s their peak race to sleep range.
- monocasa 2y agoClocks aren't the only way to race. Powering up the massive core with it's over a dozen pipelines, massive reorder buffer, etc. is another way. This is they way Apple has chose to focus on.
- oasisaimlessly 2y ago> Because clockspeed ramps power exponentially It's not exponential; P = C V^2 f, so power is something between O(f) and O(f^3), depending on how voltage scaling works. It follows that efficiency (energy per instruction) is something between O(1) and O(f^2).