Jim Keller leaves the autopilot team

[dhanson865]

He compared it to what he sees in other cars. I don't think he said anything factually incorrect.

And if you had bothered to actually watch the video you would see that they stated that the board was an nvidia.

I've worked in IT pretty much my entire adult life, I know more about computer components than he does. I do think he said some whoppers of incorrectness but I don't think it's worth my time to document them or try to correlate them as though they have any significance other than his lack of knowledge of tech.

As to actually watching the video, I did, multiple munro videos, each time they come out and make the rounds on reddit and TMC. If they'd stop making the rounds with re-edits and rehashes and summations I'd leave it alone. But repeating the same opinion over and over again doesn't make it more accurate.

The chips are Nvidia designed. Nvidia is fabless so they definitely didn't make the chips or the board. It was probably TSMC making the chips and then FoxConn or one of those folks making the board. The PCB design could have been either Nvidia or Tesla, it really doesn't matter. That's common stuff.

That's a thousand times more accurate than anything munro said about the significance of that board.

Now if Munro meant the software that goes with that board or the total system of sofware + hardware then yes it is advanced and traditional car makers in Detroit need to pay attention to what Tesla is doing. But if he wasn't including the software and was just going gaga over the hardware that board isn't that significant. Any car manufacturer can pay to have a similar board made right now. They just won't have the software to go with it and won't have it integrated into the design of the cars they are making in the factory right this second.

Maybe, just maybe I'm being pedantic and discounting Munro's praise because of the words he chose. Or more likely, he didn't phrase it well, didn't qualify his statements well, didn't setup the praise well, all because he doesn't know the tech well.

I found it slightly amusing and interesting the first time through, I have less patience to watch his statements used as a soundbite over and over in followup videos.

 

[bhzmark]

I've worked in IT pretty much my entire adult life, I know more about computer components than he does. I do think he said some whoppers of incorrectness but I don't think it's worth my time to document them or try to correlate them as though they have any significance other than his lack of knowledge of tech.

Do you disagree with his statement that tesla use of electronic components is far superior and well ahead of other car manufacturers? I found his comments plausible wrt to his comparison to car and other vehicle mfrs.

I have no basis to disagree that you know more about IT than Sandy Munro but I don’t find it relevant to the point of comparing Tesla practices to that of their competition. Sandy was persuasive on that point unless there is information correcting any material mistatements he made.

Do you know the hardware used in caddy AP? How does that compare? If you have useful info pls share in the Munro thread.

The chips are Nvidia designed. Nvidia is fabless so they definitely didn't make the chips or the board. It was probably TSMC making the chips and then FoxConn or one of those folks making the board. The PCB design could have been either Nvidia or Tesla, it really doesn't matter. That's common stuff.

Useful info . And not inconsistent with what they said on the video which was that that type of hardware is common for smaller electronics and military but not other car mfrs. that was Munros main point . He well knows that Tesla outsources many of their parts and they specifically discussed that they didn’t know the exact origin of the board he held up but mentioned that someone told them nvidia was involved.

I suggest people watch the video.

 

[mongo]

The chips are Nvidia designed. Nvidia is fabless so they definitely didn't make the chips or the board. It was probably TSMC making the chips and then FoxConn or one of those folks making the board. The PCB design could have been either Nvidia or Tesla, it really doesn't matter. That's common stuff.

Made the chips/ boards in the sense of directly or via contractors produced them. PX2/ Jetson/ GTX Titan are considered Nvidia designs. I.e. was Nvidia the lead role or not.

If Nvidia lead the board development, then that will make it harder for Tesla if they go with a different chip in the future. If Tesla lead the development, then the knowledge is, and stays, in house.

Still not sure what Keller was doing there. His specialty is not really what Tesla is trying to do. Tesla wants low powered ASICs like inference chips without relying on Nvidia. This is way more work than it sounds like though. The chips themselves are rather simple (kind of the whole point) but you need to make all the software you use target them. For this you need more software specialists (and compiler folk at that) than pure hardware specialists. Rather than leveraging the ecosystem you're duplicating and fighting against it.

Current CPU/GPU design IS duplicating the computational units you need. GPU programming is way simpler than general CPU due to the reduced instruction set and limited I/O requirements. Load memory with the data, load other memory with the instructions, start the program and wait for completion. The overall system control is done by a standard CPU, so they aren't reinventing that. You need to know the number of compute units and their size, but that is mostly the limit of the SW impact.

Keller is into general purpose work horses. The new infotainment systems are x86 now, I believe, so maybe that was his influence. But even at that it is an Intel chip, not a custom design. So... what exactly did he do while he was there?

He was involved in the low power Apple A5 and A4, along with AMD. He knows power consumption, efficiency trade offs, memory sharing, and likely fab issues.
He is leaving to

Intel has confirmed Keller's new position there as senior vice president to oversee silicon engineering. This includes system-on-a-chip (SoC) development.

Per engaget.
Intel release

I expect in whatever the design cycle time is for new silicon, we will see a new, faster, and lower power AI processing chip.
He may also have designed/ overseen the new 48? Volt Y control chips/ architecture.

 

[Tezlanian]

Do you disagree with his statement that tesla use of electronic components is far superior and well ahead of other car manufacturers? I found his comments plausible wrt to his comparison to car and other vehicle mfrs.

Just saying it's irrelevant. The software here is what is unique and important, not the board (and chips to some extent). Any auto maker will be able to contract whoever to get hardware like this.

Made the chips/ boards in the sense of directly or via contractors produced them. PX2/ Jetson/ GTX Titan are considered Nvidia designs. I.e. was Nvidia the lead role or not.

I was being very specific. This is less specific and not in the vein of what they were discussing in the video. Since they were concerned about PCB design and creation of the board you should opt for more specificity.

If Nvidia lead the board development, then that will make it harder for Tesla if they go with a different chip in the future. If Tesla lead the development, then the knowledge is, and stays, in house.

The toolchain and targeting a new architecture is the bigger deal. Who, specifically, created the PCB doesn't matter much.

Current CPU/GPU design IS duplicating the computational units you need. GPU programming is way simpler than general CPU due to the reduced instruction set and limited I/O requirements. Load memory with the data, load other memory with the instructions, start the program and wait for completion. The overall system control is done by a standard CPU, so they aren't reinventing that. You need to know the number of compute units and their size, but that is mostly the limit of the SW impact.

This is inaccurate. GPUs are fairly complex ISAs (perhaps you're only comparing x86-64?). The whole point of these simplified tensor focused you see from Nvidia (and others) is to reduce and strip precision to more efficiently use transistors while maintaining compatibility with the existing toolchain.

I believe you're experience perspective bias against the reality that the main motherboard with CPU (and now on die memory controllers) exchanging data with its system memory and the GPU exchanging data with its GPU memory are identical in nature. When contrasting the two, this is the accurate scenario. The scenario where GPU requests data from system memory to execute your kernel is another.

GPUs are actually far hungrier in their I/O requirements and have been since their inception. That's why you always see GPUs with memory generations ahead of system memory to keep the unit fed.

He was involved in the low power Apple A5 and A4, along with AMD. He knows power consumption, efficiency trade offs, memory sharing, and likely fab issues.
He is leaving to
Per engaget.
Intel release

I expect in whatever the design cycle time is for new silicon, we will see a new, faster, and lower power AI processing chip.
He may also have designed/ overseen the new 48? Volt Y control chips/ architecture.

The chips Tesla wants are far more simple than what he usually works on though. Also, when Lattner left it was mentioned what he accomplished while there. It's strange they did not for Keller. I think my question is valid and suspect he just spent time fighting politics and eventually left.

 

[GeorgeSymonds]

I've heard the Director responsible for EMEA is also leaving. Seems to be a constant stream of people leaving in top jobs, maybe its normal for tech companies but people like Jony Ive is still at Apple.

 

[pkodali]

I've heard the Director responsible for EMEA is also leaving. Seems to be a constant stream of people leaving in top jobs, maybe its normal for tech companies but people like Jony Ive is still at Apple.

I would say that Franz is the equivalent of johnny Ive at tesla