Conversation

These intelligence augmentation systems all operate so differently at 1s latency vs. 100ms latency vs. 10ms latency!

Quote Tweet

· Aug 30, 2018

I added an automatic refresh rate controller to pix2pix. Now the UI runs at 60 fps. I can draw things without any hiccups. github.com/keijiro/Pix2Pix

0:18

29.3K views

Omar Rizwan

@rsnous

Aug 31, 2018

I'm sure you've seen this, but I immediately thought of youtube.com/watch?v=vOvQCP (which I think I originally got from

@danluu

) A video I've come back to over and over since I first saw it

Replying to

and

100%! I was part of a project trying to cut one frame of latency out of iOS's touch response. Cutting just a few ms down required changes top-to-bottom in the OS! Alt story: iOS touch code had a fun rule—zero dynamic allocations allowed between touch HW event and app code!

10:53 PM · Aug 31, 2018Twitter Web Client

Replying to

and

What is a dynamic allocation, and why does it slow things down?

Matt

@mattsiegel

Aug 31, 2018

it's a little like going to a crowded restaurant without having a reservation

Replying to

and

Another story came to mind you might find interesting—also probably applicable to these realtime ML systems: as of a few years ago, in some spots, iOS actually *predicts* where your touch will be one frame in the future based on past motion to squeeze out one frame of latency.

Replying to

and

What a beautiful challenge. Which part of the org set the specific requirement?

Andy Matuschak

@andy_matuschak

Sep 1, 2018

I see it as a neat convergence of top-down and bottom-up: top-down, the mandate was "60fps or death"; mid-level, the mandate was "code like embedded engineers; we'll be tracking who's slow!"; the 1-2 devs who did touch decided "that means no malloc" (among other things).

Show replies