7 ms·

I'm sorry but I'm finding it really difficult to watch this and match up what the engineer is saying to what Elon Musk is saying. For example, the engineer say

I'm sorry but I'm finding it really difficult to watch this and match up what the engineer is saying to what Elon Musk is saying. For example, the engineer says the custom ASIC does 144 TOps for 2 chips vs the NVidia drive Xavier - does 21 TOps. Okay, well yeah I expect your custom ASIC does have a nice performance advantage over the equivalent GPU. at 3.5x advantage probably seems reasonable. Cue Elon Musk: "At first seems improbable, how could it be that Tesla, who has never designed a chip before would design the best chip in the world but that is objectively what has occured. Not the best by a small margin, the best by a huge margin". Mate, it's a dot product with some memory attached, and not a single detail your half hour deep dive has gone into suggests anything other than a bog standard ASIC. "All the cars being produced right now have all the hardware necessary for full self-driving" And this is where I'm totally lost. I want to believe! But he's lied so many times now. This man is sucking the credibility out of every engineer in the room. Don't repeat the same lie twice.

tomComb 8y ago
I'm a fan of Tesla, but not Musk. You really can't believe much that comes out of his mouth. If you look at his past claims it becomes clear that he should just be ignored.
- willio58 8y ago
  Being the fan of a company while completely ignoring the ceo of said company is questionable to me.
  - tomComb 8y ago
    Tesla has great products, created and supported by thousands of people who appear to go above and beyond. Kudos to them. And kudos to Musk for the vision and the early execution, but that doesn't mean I have to like him or believe that he is the right person for this phase of the companies growth.
- davej 8y ago
  > If you look at his past claims it becomes clear that he should just be ignored. What's an example of this? His timelines are often much too aggressive for reality but he repeatedly delivers (eventually).
  - dwighttk 8y ago
    hasn't delivered full self driving (yet)
  - tomComb 8y ago
    The main example, of course, is in self-driving where he has most certainly not delivered. And he wasn't just a little off. But also his claims about being able to automate the manufacture of Model 3 turned out to be very wrong and to do terrific damage to the company - they are/were way ahead with their technology and should have taken a less risky approach to getting that tech onto the market and taking advantage of that lead. And how did he not learn from the model X experience?
    - bobsil1 8y ago
      He's more credible on physics and hardware than software.
  - kurtisc 8y ago
    >he repeatedly delivers (eventually) This is essentially unfalsifiable. In the case claims about FSD or a person on Mars are wrong then this statement can only be proven wrong when Elon Musk chooses to admit defeat - suffering a significant personal financial loss as a result - or when he retires.
dragontamer 8y ago
> For example, the engineer says the custom ASIC does 144 TOps for 2 chips vs the NVidia drive Xavier - does 21 TOps. Okay, well yeah I expect your custom ASIC does have a nice performance advantage over the equivalent GPU. at 3.5x advantage probably seems reasonable. Cue Elon Musk: Here's my issue with that. The on-chip SRAM is only 32MB, and the RAM is LPDDR4 rated at only 68GB/s. Assuming a dot-product (multiply + add) over INT8 data, that's a limitation of 2-operations per 68GB/s (that the RAM moves at). Or 136 GIOPS (Giga-integer8 operations per second). You're limited by RAM, based on what I've seen in the presentation. Unless their neural net is 32MB and fits entirely in on-chip SRAM. That seems unlikely to me...
- grandmczeb 8y ago
  Convolutions have a much higher arithmetic intensity than dot products.
  - cromwellian 8y ago
    But in terms of CNNs, it's basically dot product is it not?Can you give an example of a CNN where it isn't?
    - dragontamer 8y ago
      He has a point. There's a bit amount of data-reuse in CNNs. Hmm... it will depend on the CNN. There's probably a good neural network design that would take advantage of this architecture. IE: A well recycled convolutional layer that probably fits within the 32MB (load those weights once, use them across the whole picture). So the whole NN doesn't necessarily have to fit inside of 32MB to be useful. But at least, large portions have to fit. (say, a 128x128 tile with 20 hidden-layers is only 300kB). Recycling that portion across the 1080 x 1920 input would be relatively cheap. I herp-derped early on, there seem to be CNNs that would make good use of the architecture. Still, the memory bandwidth of that chip is very low, I'd expect GDDR6 or HBM2 to definitely be superior to the 68GBps LPDDR4 chip they put in there.
    - grandmczeb 8y ago
ehmish 8y ago
This doesn't seem wildly inconsistent with my understanding of Elon musk, in that he has a higher level of confidence of his understanding of the world than others would, and in a "you don't know what you don't know " way it seems he didn't know about custom how custom ASICs are much more efficient at specific tasks than a general purpose computer. Doesn't make him a bad person, and in fact something like that is exactly who you want leading a company trying to change the rules of the industry they're in, it's the job of the boots on the ground to exercise due diligence here.
- unityByFreedom 8y ago
  Sorry, what? Is this an auto-generated comment?
- unityByFreedom 8y ago
  I don't understand this response to the above comment. You write, > it seems he didn't know about custom how custom ASICs are much more efficient at specific tasks than a general purpose computer Yet, Musk is clearly stating his new chip is state-of-the-art. He is not underselling it. > something like that is exactly who you want leading a company trying to change the rules of the industry they're in You want someone in charge who does not understand the hardware he's touting? > it's the job of the boots on the ground to exercise due diligence here Who are you referring to? Employees and investors? What, then, do you consider are the responsibilities of the CEO, if due diligence is not part of the job?
mrtron 8y ago
I think the part that Elon, and the engineer, are trying to express that you are looking past is "A dot product with some memory attached" is what they need to do low latency image inference on the camera data. They don't need a multipurpose CPU/GPU, they have a single well defined task that dominates their compute needs. They built a chip to do this that is cost and power effective, and redundant for safety.
- Traster 8y ago
  Right, but Nokia's 5G backhual infrastructure is far more capable of doing Mobile backhaul than a GPU, you don't get the head of Nokia standing up in a press conference telling everyone his team has designed the best chip in the world.
Fuzzwah 8y ago
>> how could it be that Tesla > Mate, it's a dot product with some memory attached, and not a single detail your half hour deep dive has gone into suggests anything other than a bog standard ASIC. There you go, you answered the question Musk put forward. Grats.
cma 8y ago
Is it really an ASIC? I thought they said over and over it is full custom.
- jakobegger 8y ago
  Definition from wikipedia: > An application-specific integrated circuit (ASIC /ˈeɪsɪk/) is an integrated circuit (IC) customized for a particular use, rather than intended for general-purpose use. Is there anything more "custom" than an ASIC?
  - cma 8y ago
    I suppose both are still called an ASIC. What I'm wondering is if it is gate-array/semi custom https://en.wikipedia.org/wiki/Application-specific_integrated_circuit#Gate-array_and_semi-custom_design https://en.wikipedia.org/wiki/Application-specific_integrate... or full custom: https://en.wikipedia.org/wiki/Application-specific_integrated_circuit#Full-custom_design https://en.wikipedia.org/wiki/Application-specific_integrate...
tigershark 8y ago
Apparently it’s actually less than half the performance of Nvidia solution: https://www.marketwatch.com/story/nvidia-says-tesla-inaccurate-in-self-driving-comparison-2019-04-22 https://www.marketwatch.com/story/nvidia-says-tesla-inaccura... https://www.nvidia.com/en-gb/self-driving-cars/drive-platform/hardware/ https://www.nvidia.com/en-gb/self-driving-cars/drive-platfor...

tomComb 8y ago

I'm a fan of Tesla, but not Musk. You really can't believe much that comes out of his mouth. If you look at his past claims it becomes clear that he should just be ignored.

willio58 8y ago
Being the fan of a company while completely ignoring the ceo of said company is questionable to me.
- tomComb 8y ago
  Tesla has great products, created and supported by thousands of people who appear to go above and beyond. Kudos to them. And kudos to Musk for the vision and the early execution, but that doesn't mean I have to like him or believe that he is the right person for this phase of the companies growth.
davej 8y ago
> If you look at his past claims it becomes clear that he should just be ignored. What's an example of this? His timelines are often much too aggressive for reality but he repeatedly delivers (eventually).
- dwighttk 8y ago
  hasn't delivered full self driving (yet)
- tomComb 8y ago
  The main example, of course, is in self-driving where he has most certainly not delivered. And he wasn't just a little off. But also his claims about being able to automate the manufacture of Model 3 turned out to be very wrong and to do terrific damage to the company - they are/were way ahead with their technology and should have taken a less risky approach to getting that tech onto the market and taking advantage of that lead. And how did he not learn from the model X experience?
  - bobsil1 8y ago
    He's more credible on physics and hardware than software.
- kurtisc 8y ago
  >he repeatedly delivers (eventually) This is essentially unfalsifiable. In the case claims about FSD or a person on Mars are wrong then this statement can only be proven wrong when Elon Musk chooses to admit defeat - suffering a significant personal financial loss as a result - or when he retires.

dragontamer 8y ago

> For example, the engineer says the custom ASIC does 144 TOps for 2 chips vs the NVidia drive Xavier - does 21 TOps. Okay, well yeah I expect your custom ASIC does have a nice performance advantage over the equivalent GPU. at 3.5x advantage probably seems reasonable. Cue Elon Musk: Here's my issue with that. The on-chip SRAM is only 32MB, and the RAM is LPDDR4 rated at only 68GB/s. Assuming a dot-product (multiply + add) over INT8 data, that's a limitation of 2-operations per 68GB/s (that the RAM moves at). Or 136 GIOPS (Giga-integer8 operations per second). You're limited by RAM, based on what I've seen in the presentation. Unless their neural net is 32MB and fits entirely in on-chip SRAM. That seems unlikely to me...

grandmczeb 8y ago
Convolutions have a much higher arithmetic intensity than dot products.
- cromwellian 8y ago
  But in terms of CNNs, it's basically dot product is it not?Can you give an example of a CNN where it isn't?
  - dragontamer 8y ago
    He has a point. There's a bit amount of data-reuse in CNNs. Hmm... it will depend on the CNN. There's probably a good neural network design that would take advantage of this architecture. IE: A well recycled convolutional layer that probably fits within the 32MB (load those weights once, use them across the whole picture). So the whole NN doesn't necessarily have to fit inside of 32MB to be useful. But at least, large portions have to fit. (say, a 128x128 tile with 20 hidden-layers is only 300kB). Recycling that portion across the 1080 x 1920 input would be relatively cheap. I herp-derped early on, there seem to be CNNs that would make good use of the architecture. Still, the memory bandwidth of that chip is very low, I'd expect GDDR6 or HBM2 to definitely be superior to the 68GBps LPDDR4 chip they put in there.
  - grandmczeb 8y ago

ehmish 8y ago

This doesn't seem wildly inconsistent with my understanding of Elon musk, in that he has a higher level of confidence of his understanding of the world than others would, and in a "you don't know what you don't know " way it seems he didn't know about custom how custom ASICs are much more efficient at specific tasks than a general purpose computer. Doesn't make him a bad person, and in fact something like that is exactly who you want leading a company trying to change the rules of the industry they're in, it's the job of the boots on the ground to exercise due diligence here.

unityByFreedom 8y ago
Sorry, what? Is this an auto-generated comment?
unityByFreedom 8y ago
I don't understand this response to the above comment. You write, > it seems he didn't know about custom how custom ASICs are much more efficient at specific tasks than a general purpose computer Yet, Musk is clearly stating his new chip is state-of-the-art. He is not underselling it. > something like that is exactly who you want leading a company trying to change the rules of the industry they're in You want someone in charge who does not understand the hardware he's touting? > it's the job of the boots on the ground to exercise due diligence here Who are you referring to? Employees and investors? What, then, do you consider are the responsibilities of the CEO, if due diligence is not part of the job?

mrtron 8y ago

I think the part that Elon, and the engineer, are trying to express that you are looking past is "A dot product with some memory attached" is what they need to do low latency image inference on the camera data. They don't need a multipurpose CPU/GPU, they have a single well defined task that dominates their compute needs. They built a chip to do this that is cost and power effective, and redundant for safety.

Traster 8y ago
Right, but Nokia's 5G backhual infrastructure is far more capable of doing Mobile backhaul than a GPU, you don't get the head of Nokia standing up in a press conference telling everyone his team has designed the best chip in the world.

Fuzzwah 8y ago

>> how could it be that Tesla > Mate, it's a dot product with some memory attached, and not a single detail your half hour deep dive has gone into suggests anything other than a bog standard ASIC. There you go, you answered the question Musk put forward. Grats.

cma 8y ago

Is it really an ASIC? I thought they said over and over it is full custom.

jakobegger 8y ago
Definition from wikipedia: > An application-specific integrated circuit (ASIC /ˈeɪsɪk/) is an integrated circuit (IC) customized for a particular use, rather than intended for general-purpose use. Is there anything more "custom" than an ASIC?
- cma 8y ago
  I suppose both are still called an ASIC. What I'm wondering is if it is gate-array/semi custom https://en.wikipedia.org/wiki/Application-specific_integrated_circuit#Gate-array_and_semi-custom_design https://en.wikipedia.org/wiki/Application-specific_integrate... or full custom: https://en.wikipedia.org/wiki/Application-specific_integrated_circuit#Full-custom_design https://en.wikipedia.org/wiki/Application-specific_integrate...

tigershark 8y ago

Apparently it’s actually less than half the performance of Nvidia solution: https://www.marketwatch.com/story/nvidia-says-tesla-inaccurate-in-self-driving-comparison-2019-04-22 https://www.marketwatch.com/story/nvidia-says-tesla-inaccura... https://www.nvidia.com/en-gb/self-driving-cars/drive-platform/hardware/ https://www.nvidia.com/en-gb/self-driving-cars/drive-platfor...