Tesla’s Hardware 3 and bigger neural networks

I recently wrote this article about why I’m excited for Hardware 3 (coming sometime around April).

The number of INT8 teraops (a.k.a. DNN teraops or DL teraops) Tesla is aiming for — I think something in the ballpark of 100 INT8 teraops — is a multiple of Mobileye’s EyeQ5 (24 teraops), a fraction of Nvidia’s Pegasus (320 teraops), and about same as what the Google self-driving car project said in mid-2016 it would like to have (100 teraops).

Andrej Karpathy teased that Tesla has bigger, better neural networks ready to go, but they need the new hardware before they can run them in cars.

1 Like

Nicely done. This was a really good read.

However, I dont think it will change much in the valuation of Tesla until they reliably demonstrate FSD…OR…are able to license their HW3 processor and other IP to other companies (which i think is as likely as Elon Shedding a tear into that freezes into the hell that is the stock market. ie. Nil)

Tesla has had trouble ramping up to massive output when required. So if they did launch a Robo Taxi service, do they have the infrastructure to make it happen? Even with the massive cash influx that a new higher valuation would bring in terms of raised capital. A built out of infrastructure would still take time. Growing pains that can spell doom for a company if mismanaged no matter how much capital they have.

This is why I think if they can move from just being a car maker to a services provider, its great. But ultimately, if they can license out their IP, even if the cars slow in terms of growth, the IP can still bring in revenue.



I tried asking Elon on Twitter when he thinks the Software V10 alpha build will be ready for a cross-country drive. No word yet.

Personally, I don’t know if a demo is as impressive to me as capabilities that are in ~150,000 customer cars, and that can be tested and vlogged without any influence from Tesla. We’ve had self-driving car demos since 2012. I’m a little mistrustful of them now.

Tesla has Superchargers, destination chargers, service centres, and stores in hundreds of cities. Tesla could launch in those cities today if the software (and regulations) permitted.

Part of their model is also to share revenue with customers who own HW3/HW4 Teslas. Tesla will take something like 30% of robotaxi revenue and in exchange for the other 70%, the customer will provide the capital to pay for the car and will be responsible for managing the car. Elon has described it has a hybrid between Uber and Airbnb.

So verygreen at TMC got his hands on some HW3 firmware and I participated in some analysis of it. The TMC post is here: https://teslamotorsclub.com/tmc/threads/teals-autopilot-hw3.139550/#post-3311371

I’m personally quite pleased with what they’ve discovered and where it seems to be leading. For years I’ve been hoping to hear that Tesla would choose this particular architectural approach. I think this is a great balance between complexity, risk, and performance. Hopefully there will be more reveals in time - I would love to see running code for AKNET_V9 or whatever enhanced network leads towards FSD.

1 Like