From what I understood it is an energy of 250 Wh/mile. Power does not really make sense in that case. So when the processing hardware is below 100 W of power and you drive at 60 miles per hour, it will consume abput 1,7 Wh/mile. Power consumption of the cameras comes on top of course.
2.5 Gpixels/sec is quite a lot of pixels to ingest and process for a self-driving car isn't it? Say their next model will use the following: 3x 8MP front-facing cameras at 30fps = 720Mpix/s ~= 0.75Gpix/s assume 6x 2.5MP cameras for LKA, BSD etc @ 30fps = 450Mpix/s ~=0.5Gpix/s and a single 2MP rear-view camera @ 30fps = 60Mpix/s = 0.06Gpix/s assume 2x Driver-Monitoring 2MP @ 60fps = 0.24 Gpix/s
all of that adds up to approx 1.55 Gpix/sec even if they up the fps on the front-facing cameras to 40fps, that would push the aggregate pixel rate to 1.8Gpixels/s.
there's still a fair bit of headroom left in the design.
Yes, exactly! We wonder if they will eventually use one or both of these FSD chips to connect directly to other types of sensors that can produce "image-sensor pixel-like" data. Multi-channel radar receivers perhaps?
It was said during the press conference which is available on youtube that the system has redundancy and is supposed to work even if one of the chips fail. Hence, they will probably not exploit all the 5 Gpix/s. I don't actually know if this 2.5 Gpix/s accounts only for optical sensors or also other modalities such as Radar which Tesla largely relies on. I find the Gpix/s metric weird anyway. The cameras will most likely not all run at the same resolution and with the same protocol overhead. Bits/second would make much more sense...
Interesting ...Tesla Model 3 has a power budget of 250W/mile.
ReplyDeleteWeird metric....perhaps energy/mile-travelled would make more sense.
DeleteFrom what I understood it is an energy of 250 Wh/mile. Power does not really make sense in that case. So when the processing hardware is below 100 W of power and you drive at 60 miles per hour, it will consume abput 1,7 Wh/mile. Power consumption of the cameras comes on top of course.
ReplyDeletesomeone have Kaparthy hire student
ReplyDelete/"https://arxiv.org/pdf/1812.07179.pdf"/
2.5 Gpixels/sec is quite a lot of pixels to ingest and process for a self-driving car isn't it? Say their next model will use the following:
ReplyDelete3x 8MP front-facing cameras at 30fps = 720Mpix/s ~= 0.75Gpix/s
assume 6x 2.5MP cameras for LKA, BSD etc @ 30fps = 450Mpix/s ~=0.5Gpix/s
and a single 2MP rear-view camera @ 30fps = 60Mpix/s = 0.06Gpix/s
assume 2x Driver-Monitoring 2MP @ 60fps = 0.24 Gpix/s
all of that adds up to approx 1.55 Gpix/sec
even if they up the fps on the front-facing cameras to 40fps, that would push the aggregate pixel rate to 1.8Gpixels/s.
there's still a fair bit of headroom left in the design.
Add to this that there are 2 such chips in Tesla self-driving system. So, there is 5 Gpix/s capacity in total.
DeleteYes, exactly! We wonder if they will eventually use one or both of these FSD chips to connect directly to other types of sensors that can produce "image-sensor pixel-like" data. Multi-channel radar receivers perhaps?
DeleteIt was said during the press conference which is available on youtube that the system has redundancy and is supposed to work even if one of the chips fail. Hence, they will probably not exploit all the 5 Gpix/s. I don't actually know if this 2.5 Gpix/s accounts only for optical sensors or also other modalities such as Radar which Tesla largely relies on. I find the Gpix/s metric weird anyway. The cameras will most likely not all run at the same resolution and with the same protocol overhead. Bits/second would make much more sense...
Delete