[Robotbeat]

Frankly, it'd be insane to invest in 10 or 7 or 5nm at this point for a custom design of maybe a few million units. The investment required scales insanely with reduction in feature size (Billions of dollars, potentially), and you'll never get your investment back. That effectively locks you in to an old design for several years. And the difference in unit costs or efficiency between 14 and 10nm is not big enough to matter compared to the upfront costs.

Fabbing at 14nm is the only sensible thing for Tesla to do right now, and since they're able to do it to replace their existing Nvidia chip for lower cost, it seems quite sensible.

[dragontamer]

Intel / MobilEye is making a 7nm chip for all of its customers. And part of that is because MobilEye has Ford, GM, BMW, Audi...

MobilEye is deployed in 313 car models across 27 manufacturers. They're going to have a 7nm chip next year, and are easily going to leapfrog Tesla's 14nm capabilities.

[Robotbeat]

If it's deploying so broadly over so many different models and manufacturers (with different sensor configurations, different IP, etc), then it necessarily has to be less application-specific and more general purpose. Tesla introduced its NN chip already in 2019 in actual built Model 3s already.

MobileEye plans to maybe have a 7nm chip sometime in 2020 (probably the end of the year) which might make it into year 2021 cars. So again, we're talking about ~2 years difference in deployment time for a less-tailored chip with a far more expensive development process. Maybe MobilEye can justify that (I'm not saying MobilEye is wrong to pursue 7nm), but it seems pretty clear Tesla pursued a logical plan to go with a 14nm process that is much cheaper to design for and is already available now at high quantities from multiple providers and reliable yields with zero technical risk on the chip fab side.

[dragontamer]

ASICs are all well and good when you can afford them AND when you have the volume to sustain them.

The main issue is that Tesla can't afford them. They have $2.2 Billion in cash remaining and are currently losing money at a rate of $700 Million/quarter.

Tesla's cash reserves are nose-diving and they're busy doing vanity projects like a self-driving ASIC, when their volume of cars being produced is well under 300k / year.

The volume simply isn't there yet to sustain this ASIC. 300k/year total car production, and not all of those cars would even have a self-driving capabilities installed. Tesla needs to work on expanding its Model 3 capabilities before ASICs make sense.

70,000 total vehicles produced Q1 2019. That's 280,000 vehicles this year if extrapolated over 4 quarters.

[Robotbeat]

All the more reason for them to go with the proven and far cheaper 14nm process node rather than 7nm (which Intel/MobilEye have had difficulties with as far as yields). https://www.techpowerup.com/248008/intel-at-least-5-years-be...

EDIT: Corrected below. MobilEye is using TSMC even though they're owned by Intel.

[dragontamer]

Intel/MobilEye is using TSMC 7nm for the EyeQ5 chip. There won't be any issues.

https://www.eetimes.com/document.asp?doc_id=1333990

TSMC 7nm has already deployed the iPhone A12, the iPad A12x, and AMD's Radeon VII. Its mature and ready for mass production.

-------

The problem with Tesla's 14nm chip is that their capabilities are about to be leap-frogged by a commercial-off-the-shelf solution in just a year. Note that the EyeQ5 is sampling TODAY. MobilEye is aiming at cars being deployed with the thing for 2020.

[Robotbeat]

Being potentially "leap frogged" in a couple years is only a problem if you're not planning on improving the chip in that time. Tesla is, in fact, planning on deploying an improved NN chip in 2 years (currently in development). They can use whatever process is available at that time.

And thank you for the correction about Intel/MobilEye using TSMC instead. Of course, anyone can use TSMC, so there's less vertical integration advantage for Intel/MobilEye than I had previously expected.

[hef19898]

More like all the more reason to source these chips from outside, I would argue