Currently, a race is underway around the world to build massive data centers to support the growing volume of artificial intelligence workloads, putting immense pressure on the supply of silicon, storage, and electronic components. But what if we could leverage the computing power of cars when they're not in motion to run AI inference workloads? This is precisely the somewhat outlandish idea that Tesla CEO Elon Musk proposed this week.

Thoughts of Elon Musk

Elon Musk, the world's richest man, shared his thoughts during Tesla's recent quarterly earnings call with analysts.

"One thing I've been thinking about is that if we have so many cars that might be idle… we could actually have a huge fleet of distributed inference vehicles, and then say, if they're not actively driving, we can have them doing distributed inference on a massive scale," Musk said during the call.

"Let's say the fleet is in the tens of millions, or even a hundred million, and let's say the fleet has one kilowatt of inference power, or high-performance inference power, that's a total of 100 gigawatts of inference power, and the power and cooling issues are all taken care of. That looks like a pretty significant asset."

Tesla equips its electric vehicles with AI accelerators, essential for its various autonomous driving features, including Autopilot and Full Self-Driving (FSD).

In 2014, Tesla adopted the MobileEye EyeQ3 chip, manufactured using a 40-nanometer process, to power its automatic lane keeping, automatic lane changing, automatic parking, and side collision warning functions. From 2016 to 2019, Tesla used NVIDIA GPUs to process signals from various sensors in its vehicles, including cameras, ultrasonic sensors, and radar (although Tesla has since switched to a full vision sensing system).

Since 2019, Tesla has been using its own custom-designed chips, reportedly 21 times faster than NVIDIA GPUs. The first custom chip, named HW3, features two neural network accelerators, each capable of 72 trillion operations per second. The HW3 chip powers the AI3 architecture. The HW4 chip, launched by Tesla in January 2023, is manufactured using a 7-nanometer process and offers 3 to 8 times the performance of its predecessor. It powers Tesla's AI4 architecture.

The infotainment system in Tesla's latest models also boasts powerful computing capabilities. Reportedly, it uses an AMD Ryzen processor and a dedicated AMD Navi 23 GPU, which Musk has stated can achieve 10 TFLOPS of performance, comparable to top-tier video game systems.

Source: Content compiled from hpcwired

View more at EASELINK