51–60 of 238 posts
Can future gen GPUs bridge the gap? The NVIDIA CEO seemed really confident when asked previously about the challenge that specialized AI chips posed. Is his confidence misplaced/just bluster or what?
what does he mean with regards to GPUs and CPUs "emulating"? they are computation machines doing computations. an algorithm running on one or the other isn't emulating anything in this context. running on "bare metal" is just an optimization along a certain axis. it isn't any more real of a computation. and talking about this now seems like another desperate attempt at generating hype.
By limiting the context of use, specialized hardware can perform better than CPU/GPU. I think his thinking relates strongly to ASIC crypto mining hardware https://en.bitcoin.it/wiki/ASIC .
"bare metal" is just a shorthand for the level of access to hardware that does not require a kernel/OS.
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.
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.
I think n = 8
Earlier quoted context omitted.
Yeah because all the verified neural networks made from meat travelling >100mph are legit ok to drive 100% of the time. There are no road accident deaths to speak of.
there is something fundamentally different between a machine and human. how the brain operates is not well known, so to claim our brain is a neural network made of meat is just a guess. the truth is, people have a lot of experience with other people. we can usually predict behavior of other drivers on the road, and can even predict pathological behavior or know how to handle it. that is a major point. when machines a…
Earlier quoted context omitted.
does correctness even work with deep neural networks to begin with? especially with the existence of adversarial examples, out of domain situations, etc
Adversarial examples are a thing even for humans, and yet here we are.
>By having its own silicone, Tesla can build for its own needs at its own pace. If they suddenly recognize something the hardware is lacking, they’re not waiting on someone else to build it. It’s by no means a trivial task — but if they can pull it off without breaking the bank (and Elon says it costs them “the same as the current hardware”), it could end up being a significant strength. I outright snorted at this. T…
They also published the quote about iPhone leading in performance and go-to-market because of their custom SoC efforts, and they mentioned Tesla's SoC program manager Pete Bannon, but they didn't even bother to connect the two ideas: Pete Bannon was the guy who led Apple's SoC efforts to develop the first 64 Bit ARM chip for mobile.
They also didn't mention what Elon said about now being able to run all the cameras at full resolution and full framerate.
This article could have been so much better.
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.
...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.
>By having its own silicone, Tesla can build for its own needs at its own pace. If they suddenly recognize something the hardware is lacking, they’re not waiting on someone else to build it. It’s by no means a trivial task — but if they can pull it off without breaking the bank (and Elon says it costs them “the same as the current hardware”), it could end up being a significant strength. I outright snorted at this. T…
This is not what Musk said on the conference call. Surely some bonehead reporter is writing this. Musk said the latter, precisely to go from current 200 fps speed to something like 2000 fps (I might have the numbers wrong, but it's a 10x improvement) and that they have been doing this in stealth for the last 2-3 years. They expect the cost of the new chips to be the same as that of the current nVidia lineup. That got me thinking though. What are they currently using - the Titans ?
Earlier quoted context omitted.
More data points can't hurt right? Maybe this is part of their push to move away from third party suppliers?
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…
>By having its own silicone, Tesla can build for its own needs at its own pace. If they suddenly recognize something the hardware is lacking, they’re not waiting on someone else to build it. It’s by no means a trivial task — but if they can pull it off without breaking the bank (and Elon says it costs them “the same as the current hardware”), it could end up being a significant strength. I outright snorted at this. T…
... so that they can create performance-intensive features that nobody else in the market yet has, because the off-the-shelf hardware won't allow it. Due to power consumption, cost and performance issues. And this custom hardware is being developed to solve them.
You're basically saying the same thing in different words. I mean, what else do you do with better power consumption, cost and performance other than new, better-working features? You think Tesla will sell cars to the mass market with impressive computer specs?
Apple's own custom hardware has enabled it to experiment with performance-intensive features unlike other smartphone manufacturers, Tesla is going for the same strategy.