Yesterday I wrote a few thoughts on the relationship between Tesla and AMD to develop custom chips for self-driving cars.
Sure enough, I was clearing out my inbox today and I stumbled upon this blog post by Intel CEO Brian Krzanich, about the longstanding relationship between Waymo and Intel.
The Intel blog post is dated September 18, while the CNBC report on Tesla and AMD is dated September 20, so maybe Intel beat AMD to the punch, and I’m just playing catch up.
The Krzanich post is pretty light on detail. It mainly highlights the massive number of annual automotive fatalities, and asserts that Waymo has been using Intel technology in its self-driving vehicles. No real news here.
But the Tesla-AMD announcement did capture something I neglected to mention in my previous blog post, and which I also think it is safe to assume comes into play in the Waymo-Intel relationship: custom silicon.
Off-the-shelf CPUs and GPUs are general-purpose devices that tend to do a lot of things well. GPUs are primarily built to update graphics on computer monitors, and it’s almost coincidental that this type of parallel computing happens to be really good for machine learning.
But you can imagine that building computer chips specifically for autonomous driving might yield even faster performance. The problem, of course, is how expensive it is to build a chip. The fixed costs are enormous, so they have to be amortized over millions of units of silicon.
FPGAs are a kind of intermediate solution to this problem. Not nearly as expensive to work with as custom silicon, but presumably not as fast, either.
Google (not necessarily Waymo, though) seems to be going a step farther, into full-blown chip design, with its TPUs.
That seems to be what we’re seeing both with Waymo-Intel and with Tesla-AMD, and I wouldn’t be shocked to see NVIDIA and other chipmakers go down that road as well.