Karpinski: Alternating language layers in AI frameworks prevents compiler optimizations
Stefan Karpinski · A Fresh Approach to Technical Computing // Viral Shah & Stefan Karpinski, Julia Computing · Sep 28, 2017 · at 6:39
Stefan Karpinski, co-creator and CTO of Julia Computing, explains performance limitations in multi-language AI frameworks like TensorFlow and PyTorch.
0:00 / 0:24exact quote · 24.8s
▶ Watch the full episode on YouTube → 720p mp4 · rendered on demand · StarZero watermark
“They have what I've, I would describe as a sandwich problem, which is that you end up sandwiching a lot of system code with user code, and then, like, adding more and more layers of that, and as you've sandwiched, like, you know, seven or eight layers of that, you know, you might be doing a really trivial operation, but it can't be collapsed because every time you hit that barrier where you switch languages, there is no possibility for optimization.”
quote is from the automated transcript, cleaned for reading: filler sounds and stutters are removed, nothing is rephrased. names can be misheard (the analysis reads context, assessments check outside sources). how →
More from Stefan Karpinski
Karpinski: R produces the best plots and leads statistical analysis
“I think it makes the best plots and it's, you know, the undisputed king in, in statistical analysis.”
Karpinski: Julia surpassed NumPy in contributors within one year
“Even when we were one year into being a public project, we already had more contributors and developers than NumPy, and I could not understand how that was possible, because I was like, we're this very young project, this was, you know, back in 2013”
Karpinski: Julia's real-world execution speed stays within 2x of C
“The real life, life experience of what people have done, you know, in like, you know, years of programming in Julia now, is that they tend to see that Julia is within one to two of C.”
Karpinski: Automatic dynamic code translation fails due to edge cases
“The trouble is when you get into automatic translation, it's always the corner cases that kill you. And the corner cases that kill you with automatic translation are actually exactly the reasons why you can't just sit down and be like, oh, let's just make Pyth…”