24 ms·
> Histortically, the main value of GIL removal proposals and implementations has been to spur the core team to speed up single core codes. While I'm all for i
by godelski 3y ago
> Histortically, the main value of GIL removal proposals and implementations has been to spur the core team to speed up single core codes.
While I'm all for increasing single threaded performance[side rant] is this really a good argument? My understanding is that Moore's Law only currently exists because of the existence of more cores. Clock frequency has stagnated since 2004[0]. I mean feature size (Moore's Law) still exists[1] but IPS/core seems significantly flatter than IPS[2]. Aren't most of our performance gains from mutlicore? And as we push more in this direction (e.g. GPU taking over compute) doesn't this make this even worse for python? (Yes I understand that you can make C/CUDA calls but doesn't the GIL make problems here as well as prevent a native python solution?)
[side rant] Still frustrated by things like that difference in speeds between math.sqrt(2.) (~0.002s/10k), np.sqrt(2.) (~0.010/10k), torch.sqrt(torch.tensor([2.])) (~0.039/10k). I know there's more going on in capabilities, but it is of course frustrating by the large differences.
[0] http://cpudb.stanford.edu/visualize/clock_frequency http://cpudb.stanford.edu/visualize/clock_frequency
[1] http://cpudb.stanford.edu/visualize/technology_scaling http://cpudb.stanford.edu/visualize/technology_scaling
[2] https://en.wikipedia.org/wiki/Instructions_per_second#Millions_of_instructions_per_second_(MIPS) https://en.wikipedia.org/wiki/Instructions_per_second#Millio...
- aragilar 3y agoThat difference (at least in the math->numpy case) is entirely down to conversions (or at least it was when I profiled it ~7 years ago). If you're careful with types (i.e. avoiding converting from float to numpy scalar values), then the difference disappears.