Neural Networks

[mongo]

TOPS are not TFLOPS

Trillions of (usually 8-bit integer for NN) Operations Per Second vs Trillions of Floating-point Operations Per Second

I'm sure he meant TOPS not TFLOPS and just got hung up on the more common term. OR should have meant it

Yah, I know, TFOPS are good for training though... For NN operation, perhaps MACs from DSP land would be more useful? Unless all ops take the same time...

 

[chillaban]

Yah, I know, TFOPS are good for training though... For NN operation, perhaps MACs from DSP land would be more useful? Unless all ops take the same time...

That’s basically what they implied their accelerator is.... a matrix multiply accelerator, performing low precision FO MACs.

 

[strangecosmos]

@jimmy_d What do you think about Karpathy’s comment that “as you make the networks bigger by adding more neurons, the accuracy of all their predictions increases with the added capacity”? Did he really mean neurons, or did he mean weights? Could we see more neuron layers in a future neural network on Hardware 3?

 

[scaesare]

... then again I've never been wrong so...

Aww how cute... you have your very own version of The Liar's Paradox!:

 

[scaesare]

Also you can't just dump random images from Google Earth into the captcha service and train it using the general public. Captcha needs to know there's a sign in that image and which tiles are covered by the sign in order for their service to work. Implying that this metadata is already there labeled manually.

I've always assumed these are images that the NN thought contained signs (or numbers, etc...) and then distributed to a statistically significant number of unwitting captcha validators...

 

[mongo]

@jimmy_d What do you think about Karpathy’s comment that “as you make the networks bigger by adding more neurons, the accuracy of all their predictions increases with the added capacity”? Did he really mean neurons, or did he mean weights? Could we see more neuron layers in a future neural network on Hardware 3?

Not Jimmy, but I'd say neurons, weights without neurons means pulling more values from the previous layer. It increases the area the neuron is looking at, but does not increase the number of different things the layer itself can detect.

The SW 2.0 talk Karpathy gave may be the NN setup they are going with. Although SW 2.0 could reconfigure it to be something else...

 

[BioSehnsucht]

Not Jimmy, but I'd say neurons, weights without neurons means pulling more values from the previous layer. It increases the area the neuron is looking at, but does not increase the number of different things the layer itself can detect.

The SW 2.0 talk Karpathy gave may be the NN setup they are going with. Although SW 2.0 could reconfigure it to be something else...

Or maybe SW 2.0 is the machine that trains the machine?

 

[mongo]

Or maybe SW 2.0 is the machine that trains the machine?

Which is the machine being trained..

So then aliens dreadnought will be the machine that builds the machine that trains the machine that is the machine that machines the machines that Elon built...

 

[BioSehnsucht]

Which is the machine being trained..

So then aliens dreadnought will be the machine that builds the machine that trains the machine that is the machine that machines the machines that Elon built...

Careful, or we'll have analysts worrying about it ending up like this :

 

[Sharps97]

We sure can. Eyeq4 is currently in production in multiple cars and has 2.5 TFLOPS while running on 3 Watts. It processes a more vast variety of NN compared to anything Tesla has with far better accuracy on 8+ cameras at 36 frames per second.

There are some that might suggest that Mobileye is being slightly... disingenuous with their spec quoting:
https://www.eetimes.com/document.asp?doc_id=1332687

Best guess is that the Eyeq4 "platform" is > 3 watts and that is what Tesla is referring to on a comparative basis (platform power). They've given no real specifics here, so a meaningful comparison remains impossible on a hardware level. That said, compared to your typical ARM core, NVidia's Xavier is a computing monster and I doubt that the Eyeq4 platform can come close to keeping up with it.

If you look at the block diagram for Xavier, the NN is only a small part of the overall package:
Tegra Xavier - Nvidia - WikiChip

So basically, if you just consider the NN part of the die only, then they (Nvidia, Mobileye, and Tesla) are all probably low single digit watt parts assuming a manufacturing node of 14 nm or smaller.

From a software perspective, we know that Mobileye advertises more capability than Tesla does right now. That said, it does not appear to translate into an actual product that exceeds what Cruise is delivering (today), and that appears more dependent on high resolution mapping than actual image recognition. So even if the paper capabilities are better, the delivered products are simply... different with one not meaningfully better than the other (right now).

Elon also indicated that AP 9.x right now does not fully utilize the compute resources of HW 2.0. However, the next iteration apparently will (but no details have been provided by Elon or Tesla as to what exactly those expected capabilities are planned to be). It is great that there are those on this forum who have provided some insight based on 3rd party evaluations, but that is not the same as actually having the full picture presented by the company. For all we know, the "super NN" waiting in the wings will do everything that Eyeq5 will do - but there is no way to prove / disprove based on actually available data (i.e., what was recently assessed is unlikely to even be the latest / greatest version that Telsa is working on).

Ultimately, I think it is too early to call victory (or failure) based on what we have high confidence knowledge on. We know that it is possible for Tesla to leverage existing hardware to do autonomous lange changes, but there are undefined edge cases that must be resolved (and no specifics on exactly what those are or how often they occur). But it seems likely that additional minor changes are possible to improve the product until version 3.0 of the hardware comes out. Right now, both companies are talking about what they eventually hope to get to - and debating speculation vs. speculation isn't something that will be a settled thing any time soon.

 

[strangecosmos]

Soon every thread in this entire forum is going to be Bladerskb pulling people into off-topic arguments about Mobileye. Rename it the "Mobileye vs. Tesla" forum

 

[S4WRXTTCS]

So basically, if you just consider the NN part of the die only, then they (Nvidia, Mobileye, and Tesla) are all probably low single digit watt parts assuming a manufacturing node of 14 nm or smaller.

Just a quick clarification.

The two Deep Learning Accelerators that's shown on the die picture are not where the bulk of the DL interference is going to run.

What they're really good at is efficient interference. Like lets say you built a robot that you wanted to wake up when some event happened. Instead of using the integrated GPU+Tensor Cores you'd use the one of the Deep Learning accelerators.

I have a Jetson Xavier in my Rover project, and this is what I plan on using it for.

The other thing that's really interesting about Nvidia is they tend to rapidly improve acceleration with each release of TensorRt. So the current numbers aren't really a good indication of the limits. With the Nvidia TX2 in particular there were some significant speeds up in inference speed due to more efficient acceleration.

I don't think Tesla is going to move away from using Nvidia's SOC's. Instead they're just going to have a dedicated neural net accelerator attached to them.

I can't wait to find out what SOC's the HW3 uses.

Nvidia created a pretty good presentation of the Xavier SOC in case anyone in interested.

dusty-nv/jetson-presentations

 

[J1mbo]

Peter Bannon

Hi, this is Pete Bannon. The Hardware 3 design is continuing to move along. Over the last quarter, we've completed qualification of the silicon, qualification of the board. We started the manufacturing line, in qualification of the manufacturing line. We've been validating the provisioning flows in the factor. We built test versions of Model S, X and 3 in the factory to validate all the fit and finish of the parts and all the provisioning flows.

So we still have a lot of work to do. And the team is doing a great job, and we're still on track to have it ready to go by the end of Q1.

This tells me that full-fat HW3 is more than a simple replacement of the AP box.

Am willing to bet HW3 will bring the selfie cam to S and X, maybe adding IR illumination for night driving. But IMO that does not require a whole test build. Corner radar? Changes to the pillar cams to stop them from misting up?

 

[lunitiks]

This tells me that full-fat HW3 is more than a simple replacement of the AP box.

Am willing to bet HW3 will bring the selfie cam to S and X, maybe adding IR illumination for night driving. But IMO that does not require a whole test build. Corner radar? Changes to the pillar cams to stop them from misting up?

Interesting point on Bannon’s comment

 

[Ulmo]

This tells me that full-fat HW3 is more than a simple replacement of the AP box.

Am willing to bet HW3 will bring the selfie cam to S and X, maybe adding IR illumination for night driving. But IMO that does not require a whole test build. Corner radar? Changes to the pillar cams to stop them from misting up?

It could be just the fit and finish of the computational hardware or some of the brackets and trim instructions necessary to replace it, since this is going to be a properly instructed mass-replacement effort that should be given the best chance of good results on a mass scale. The fine level of detail needed for such instructions can often have more refined information per word than there are words. (One would hope these aren't Ikea instructions written by translators, marketing, lawyers, corporate commitee, and ESL students, in backwards sequential order and thus also out of date, wrong, incomprehensible, incorrect, meaningless, fake, and confusing.)

 

[croman]

It could be just the fit and finish of the computational hardware or some of the brackets and trim instructions necessary to replace it, since this is going to be a properly instructed mass-replacement effort that should be given the best chance of good results on a mass scale. The fine level of detail needed for such instructions can often have more refined information per word than there are words. (One would hope these aren't Ikea instructions written by translators, marketing, lawyers, corporate commitee, and ESL students, in backwards sequential order and thus also out of date, wrong, incomprehensible, incorrect, meaningless, fake, and confusing.)

This. It's just the trim that is the subject of the sentence. Building the cars means using this trim kit instead of hw2.5. perhaps it's also wiring harness too. If it's actual hardware then Musk lied during his call and the SEC will tear into him again.

 

[BioSehnsucht]

I don't doubt the sensor suite will evolve (perhaps better lenses, or wider dynamic range, etc), but they are pretty solidly comitted to FSD working on AP2/2.5 vehicles with the AP3 brain upgrade.

 

[Ulmo]

I don't doubt the sensor suite will evolve (perhaps better lenses, or wider dynamic range, etc), but they are pretty solidly comitted to FSD working on AP2/2.5 vehicles with the AP3 brain upgrade.

Yes; there's no reason why not to improve the whole thing for new vehicles while making the ones promised to work work as promised, which is exactly what they said they would do, and the stated intentional general approach Tesla has used for a decade with respect to vehicle improvements (and which often starts many arguments on TMC of people mad at Tesla for improving newer vehicles).

 

[J1mbo]

Whilst I agree with the above points, Tesla upgrades components "every week". I am pretty sure they don't built test cars for every upgrade.

Also, the AP3 box is supposed to be a 30 min swap at a Service Center. That means the physical mounting points and cables must be in the same place on the new unit. No need for a pre-build if the assembly step is going to be the same.

I guess it is also possible that they will have two SKUs - one for new build, and one for retrofit.

 

[mongo]

This tells me that full-fat HW3 is more than a simple replacement of the AP box.

Am willing to bet HW3 will bring the selfie cam to S and X, maybe adding IR illumination for night driving. But IMO that does not require a whole test build. Corner radar? Changes to the pillar cams to stop them from misting up?

S/X and 3 AP HW currently have different mounting and cooling features.
My take: The selfie cam is part of Tesla Network which will only be on cars going forward. So 3 and later build S/X. Older versions will have FSD but require a driver in the car. Future cars may all have water cooled AP HW, but AP 2.0 is air cooled.