7 ms·

The benefit they point out is 200fps vs. 2000fps, but is that actually that useful? In the post of OpenAO's robot, they said that improving the response times g

The benefit they point out is 200fps vs. 2000fps, but is that actually that useful? In the post of OpenAO's robot, they said that improving the response times gave no significant gains. Only thing I can imagine would be allowances for wider and deeper neural nets.

fhood 8y ago
More data points can't hurt right? Maybe this is part of their push to move away from third party suppliers?
- acdha 8y ago
  More data points can hurt: it may just waste time processing data which doesn't improve results but it can actually make things harder if you're just adding more noise which has to be filtered out but there's no corresponding improvement in signal. As a simple example in a different domain, years back I found that the best way to improve Tesseract's OCR accuracy was to ensure that I didn't feed it images at more than 150dpi because it would sometimes misrecognize dust, paper texture, etc. as characters.
  - qaq 8y ago
    My understanding is big problem is stationary objects which mostly get filtered out. With more processing power you might be able to do something about that.
    - acdha 8y ago
      My question is where the break-even point is between adding more resolution / frames-per-second versus performing more expensive analysis. Given that the human visual system is not especially high resolution or speed, I would bet that they're already at the point of diminishing returns on the sensor side and the big wins would come from more expensive processing techniques, which https://news.ycombinator.com/item?id=17671843 https://news.ycombinator.com/item?id=17671843 supports.
chomp 8y ago
>2000fps will provide a lot more wiggle room to do higher level computation on each frame. Yeah I think this is it. My best guess, it's 200fps vs 2000fps for the same workload. As Tesla's data model gets larger and its computations get more complex, that 200fps will start to sag. 2000fps will provide a lot more wiggle room to do higher level computation on each frame.
- bonestamp2 8y ago
  This is exactly right. They didn't mention it in the article, but I was on the call and Elon stated that one of the main benefits was that they could now process all of the camera images at full resolution and full framerate (implying they can't do that with the GPU).
AbrahamParangi 8y ago
I bet the performance improvement with higher sampling frequencies is related to the percent of the input signal represented in the sampling (https://en.wikipedia.org/wiki/Nyquist%E2%80%93Shannon_sampling_theorem https://en.wikipedia.org/wiki/Nyquist%E2%80%93Shannon_sampli...)
reacweb 8y ago
Maybe that chip was not important in itself but was just a good opportunity to improve in house expertise in chip development. Having expertise gives a lot of power during negociations with providers.
pasta 8y ago
Let's say you drive 130km/h (36m/s). 36 / 200 frames = 18cm travel distance between each calculation. So you either can calculate much more each 18cm or calculate shorter distances. I think a shorter distance is not very useful but calculating more might be.
- auferstehung 8y ago
  *n cameras I think n = 8
- baybal2 8y ago
  One thing high FPS are good for are optical flow mappers. Think of estimating relative motion of things in video.
gvb 8y ago
> Only thing I can imagine would be allowances for wider and deeper neural nets. ...or more cameras, with better frame rates over a lot of cameras. My guess is that the fps rate is over all the cameras, so 10 cameras means going from 20fps per camera to 200fps per camera.
visarga 8y ago
I think latency makes a lot of difference when an autonomous system is interacting with the world. Take a look at this video from 2009 where they use a high speed vision system (1000fps!). The things it could do are probably unequaled even today. https://www.youtube.com/watch?v=-KxjVlaLBmk https://www.youtube.com/watch?v=-KxjVlaLBmk The higher the speed of the system, the less it has to predict in the future and better can adapt to unforeseen situations. It's basically using the world as a model, for free.
- mannykannot 8y ago
  That is impressive, though, in the case of cars, is the minimum prediction not determined by the physics of steering and braking?

fhood 8y ago

More data points can't hurt right? Maybe this is part of their push to move away from third party suppliers?

acdha 8y ago
More data points can hurt: it may just waste time processing data which doesn't improve results but it can actually make things harder if you're just adding more noise which has to be filtered out but there's no corresponding improvement in signal. As a simple example in a different domain, years back I found that the best way to improve Tesseract's OCR accuracy was to ensure that I didn't feed it images at more than 150dpi because it would sometimes misrecognize dust, paper texture, etc. as characters.
- qaq 8y ago
  My understanding is big problem is stationary objects which mostly get filtered out. With more processing power you might be able to do something about that.
  - acdha 8y ago
    My question is where the break-even point is between adding more resolution / frames-per-second versus performing more expensive analysis. Given that the human visual system is not especially high resolution or speed, I would bet that they're already at the point of diminishing returns on the sensor side and the big wins would come from more expensive processing techniques, which https://news.ycombinator.com/item?id=17671843 https://news.ycombinator.com/item?id=17671843 supports.

chomp 8y ago

>2000fps will provide a lot more wiggle room to do higher level computation on each frame. Yeah I think this is it. My best guess, it's 200fps vs 2000fps for the same workload. As Tesla's data model gets larger and its computations get more complex, that 200fps will start to sag. 2000fps will provide a lot more wiggle room to do higher level computation on each frame.

bonestamp2 8y ago
This is exactly right. They didn't mention it in the article, but I was on the call and Elon stated that one of the main benefits was that they could now process all of the camera images at full resolution and full framerate (implying they can't do that with the GPU).

AbrahamParangi 8y ago

I bet the performance improvement with higher sampling frequencies is related to the percent of the input signal represented in the sampling (https://en.wikipedia.org/wiki/Nyquist%E2%80%93Shannon_sampling_theorem https://en.wikipedia.org/wiki/Nyquist%E2%80%93Shannon_sampli...)

reacweb 8y ago

Maybe that chip was not important in itself but was just a good opportunity to improve in house expertise in chip development. Having expertise gives a lot of power during negociations with providers.

pasta 8y ago

Let's say you drive 130km/h (36m/s). 36 / 200 frames = 18cm travel distance between each calculation. So you either can calculate much more each 18cm or calculate shorter distances. I think a shorter distance is not very useful but calculating more might be.

auferstehung 8y ago
*n cameras I think n = 8
baybal2 8y ago
One thing high FPS are good for are optical flow mappers. Think of estimating relative motion of things in video.

gvb 8y ago

> Only thing I can imagine would be allowances for wider and deeper neural nets. ...or more cameras, with better frame rates over a lot of cameras. My guess is that the fps rate is over all the cameras, so 10 cameras means going from 20fps per camera to 200fps per camera.

visarga 8y ago

I think latency makes a lot of difference when an autonomous system is interacting with the world. Take a look at this video from 2009 where they use a high speed vision system (1000fps!). The things it could do are probably unequaled even today. https://www.youtube.com/watch?v=-KxjVlaLBmk https://www.youtube.com/watch?v=-KxjVlaLBmk The higher the speed of the system, the less it has to predict in the future and better can adapt to unforeseen situations. It's basically using the world as a model, for free.

mannykannot 8y ago
That is impressive, though, in the case of cars, is the minimum prediction not determined by the physics of steering and braking?