4 ms·

There's more to it than feature size. Other companies have more general-purpose chips. Tesla basically built an ASIC for self-driving.

dragontamer 8y ago
NVidia is the only chip that seems a bit heavy on GPU features. But the Volta architecture has tensor processors with full FP16 and INT8 support, so it seems useful for self-driving as far as I can tell. MobilEye EyeQ4 and EyeQ5 has no other purpose than self-driving. EyeQ5 is going to be 7nm and is launching 2020. A full ASIC for self-driving as well. And no one knows what the heck Google / Waymo is doing. All we know is that its working, today, in Phoenix Arizona, where anyone can use Google / Waymo's taxi service TODAY. We also know that Google employes the same researchers who made TPUs that powered AlphaGo. They're very secretive about their hardware (for good reason!), but they're the elephant in the room with regards to self-driving. https://waymo.com/apply/ https://waymo.com/apply/
- Judgmentality 8y ago
  > And no one knows what the heck Google / Waymo is doing. All we know is that its working, today, in Phoenix Arizona, where anyone can use Google / Waymo's taxi service TODAY. This is false. There are, to the best of anyone's knowledge, a handful of people in the Waymo One program (I'm fairly confident it's less than 10). Everyone in the Early Rider program is still under NDA. In order to be accepted to either, you have to apply. The application process being open to the public does not mean the service is public, anymore than the application process to Google being public means anyone can work at Google. As to regards to how well it's working...they are definitely in the lead, and they do really have cars driving around, but all of them are still with safety drivers. Until they pull the safety drivers you can't really call them driverless and it shows a lack of faith in their product because it's not ready yet to actually be driverless. Waymo lost their goodwill when they said they'd have the service you mentioned in 2018, basically pulled a bullshit PR event in order to "fulfill" their promise. It's pretty obvious though it was just so whoever was in charge could get some bonus attached to the timeline, very similar to how Google seems to launch products already - they do it so people can get promoted and then all the important people leave before the product is useful to anyone. It's no secret that almost the entire original team has already left and most of them have started competing against Waymo, so it seems they are confident they can take on Waymo and win - and they'd know better than anyone because they built Waymo. And Waymo has since gone silent on when they'll expand their program, or pull the driver from the car, or pretty much anything. I agree it's a good idea for them to play their cards close to their chest, but compared to their past posturing that's a sign that it's not around the corner.
  - dragontamer 8y ago
    My point is that Waymo has taken large and public steps forward. People have seen the Waymo cars (and have to deal with them) in Phoenix. No one has ever seen a "Full self driving" car from Tesla, despite the fact that hundreds-of-thousands of people have paid $5000 for the promise of future hardware to be installed onto their car. My point is: other companies are ahead, and have demonstrated capabilities above and beyond Tesla's 14nm chip team.
    - tim333 8y ago
      I'm not convinced one way or the other which is ahead out of Tesla and Waymo. Waymo have low interventions per mile but don't seem able to deal with heavy traffic for merging or turning. Tesla have had issues driving into things but are out there in heavy traffic.
- Robotbeat 8y ago
  Launching in 2020 means two years (starting in 2018) of using the more expensive and slower Nvidia chip. By using the 14nm process now, Tesla paid far less on upfront dev costs (we're talking tens of millions versus billions) and saved on unit costs right out the gate, putting them in better position for smaller process sizes in the future. Volta architecture is much more general purpose than Tesla's. If you're doing a custom ASIC hyper-focused on your application and you're not selling hundreds of millions of them, it doesn't make sense to lock yourself into paying of dev costs for cutting edge fabrication techniques with risk of yield problems and questionable per-unit economics.
  - dragontamer 8y ago
    > Launching in 2020 means two years (starting in 2018) of using the more expensive and slower Nvidia chip. You mean a cheaper and more industry established MobilEye EyeQ5 chip, with LiDAR capabilities, Tensor Units and all that good stuff. ------------ Let me ask you this: how many vehicles do you think Tesla will make in 2018 and 2019? A million? Two million? Do remember, 2018 Tesla made only 244,920 cars. 2019 Q1 only had 70,000 cars, so Tesla is currently producing cars at a rate under 300,000 cars/year (maybe growing to 350,000 by the end of the year). Maybe Gigafactory 3 comes up in early 2020, and Tesla then makes 500k cars for 2020. So... just about a million cars between 2018 to 2020, and I'm being quite generous on Tesla's growth potential here. Remember: not all of those cars having FSD or autopilot installed. I'm not seeing how this 14nm chip is going to be a financially sound move for Tesla. They just don't make enough cars and they weren't watching out on industry standard chip upgrades.
    - Robotbeat 8y ago
      14nm is a sound move for Tesla exactly because their volumes are lower. All their cars will have the new chip installed as it's part of the safety features as well (and gives them margin whenever someone decides to do an upgrade on their existing vehicle).
      - hef19898 8y ago
        With these low volumes the sound move would be to buy chips from someone else. Just like Aston Martin is sourcing engines from third parties. Heck, even Mercedes is using Renault engines and BMW used Peugeot engines in Minis for years. And engines are a core competency of both, BMW and Mercedes. Chips for Tesla, not so much.

dragontamer 8y ago

NVidia is the only chip that seems a bit heavy on GPU features. But the Volta architecture has tensor processors with full FP16 and INT8 support, so it seems useful for self-driving as far as I can tell. MobilEye EyeQ4 and EyeQ5 has no other purpose than self-driving. EyeQ5 is going to be 7nm and is launching 2020. A full ASIC for self-driving as well. And no one knows what the heck Google / Waymo is doing. All we know is that its working, today, in Phoenix Arizona, where anyone can use Google / Waymo's taxi service TODAY. We also know that Google employes the same researchers who made TPUs that powered AlphaGo. They're very secretive about their hardware (for good reason!), but they're the elephant in the room with regards to self-driving. https://waymo.com/apply/ https://waymo.com/apply/

Judgmentality 8y ago
> And no one knows what the heck Google / Waymo is doing. All we know is that its working, today, in Phoenix Arizona, where anyone can use Google / Waymo's taxi service TODAY. This is false. There are, to the best of anyone's knowledge, a handful of people in the Waymo One program (I'm fairly confident it's less than 10). Everyone in the Early Rider program is still under NDA. In order to be accepted to either, you have to apply. The application process being open to the public does not mean the service is public, anymore than the application process to Google being public means anyone can work at Google. As to regards to how well it's working...they are definitely in the lead, and they do really have cars driving around, but all of them are still with safety drivers. Until they pull the safety drivers you can't really call them driverless and it shows a lack of faith in their product because it's not ready yet to actually be driverless. Waymo lost their goodwill when they said they'd have the service you mentioned in 2018, basically pulled a bullshit PR event in order to "fulfill" their promise. It's pretty obvious though it was just so whoever was in charge could get some bonus attached to the timeline, very similar to how Google seems to launch products already - they do it so people can get promoted and then all the important people leave before the product is useful to anyone. It's no secret that almost the entire original team has already left and most of them have started competing against Waymo, so it seems they are confident they can take on Waymo and win - and they'd know better than anyone because they built Waymo. And Waymo has since gone silent on when they'll expand their program, or pull the driver from the car, or pretty much anything. I agree it's a good idea for them to play their cards close to their chest, but compared to their past posturing that's a sign that it's not around the corner.
- dragontamer 8y ago
  My point is that Waymo has taken large and public steps forward. People have seen the Waymo cars (and have to deal with them) in Phoenix. No one has ever seen a "Full self driving" car from Tesla, despite the fact that hundreds-of-thousands of people have paid $5000 for the promise of future hardware to be installed onto their car. My point is: other companies are ahead, and have demonstrated capabilities above and beyond Tesla's 14nm chip team.
  - tim333 8y ago
    I'm not convinced one way or the other which is ahead out of Tesla and Waymo. Waymo have low interventions per mile but don't seem able to deal with heavy traffic for merging or turning. Tesla have had issues driving into things but are out there in heavy traffic.
Robotbeat 8y ago
Launching in 2020 means two years (starting in 2018) of using the more expensive and slower Nvidia chip. By using the 14nm process now, Tesla paid far less on upfront dev costs (we're talking tens of millions versus billions) and saved on unit costs right out the gate, putting them in better position for smaller process sizes in the future. Volta architecture is much more general purpose than Tesla's. If you're doing a custom ASIC hyper-focused on your application and you're not selling hundreds of millions of them, it doesn't make sense to lock yourself into paying of dev costs for cutting edge fabrication techniques with risk of yield problems and questionable per-unit economics.
- dragontamer 8y ago
  > Launching in 2020 means two years (starting in 2018) of using the more expensive and slower Nvidia chip. You mean a cheaper and more industry established MobilEye EyeQ5 chip, with LiDAR capabilities, Tensor Units and all that good stuff. ------------ Let me ask you this: how many vehicles do you think Tesla will make in 2018 and 2019? A million? Two million? Do remember, 2018 Tesla made only 244,920 cars. 2019 Q1 only had 70,000 cars, so Tesla is currently producing cars at a rate under 300,000 cars/year (maybe growing to 350,000 by the end of the year). Maybe Gigafactory 3 comes up in early 2020, and Tesla then makes 500k cars for 2020. So... just about a million cars between 2018 to 2020, and I'm being quite generous on Tesla's growth potential here. Remember: not all of those cars having FSD or autopilot installed. I'm not seeing how this 14nm chip is going to be a financially sound move for Tesla. They just don't make enough cars and they weren't watching out on industry standard chip upgrades.
  - Robotbeat 8y ago
    14nm is a sound move for Tesla exactly because their volumes are lower. All their cars will have the new chip installed as it's part of the safety features as well (and gives them margin whenever someone decides to do an upgrade on their existing vehicle).
    - hef19898 8y ago
      With these low volumes the sound move would be to buy chips from someone else. Just like Aston Martin is sourcing engines from third parties. Heck, even Mercedes is using Renault engines and BMW used Peugeot engines in Minis for years. And engines are a core competency of both, BMW and Mercedes. Chips for Tesla, not so much.