Isn't Tesla's FSD computer best compared to Google's TPU? Both use integer operations instead of floating point and are optimized for matrix operations.

It's a great product and genuinely impressive, but you're comparing specialist vs generalist hardware. Half an order of magnitude improvements are table stakes, otherwise it's crazy not to just buy off the shelf.

My 2070 Super does 9 TFLOPs but it's paired with an 3900X which barely does 1... is my graphics card an order of magnitude better than my CPU?

It's cool that Tesla made an ASIC, I guess.

I think more relevant though, would be Tesla's hype around "Autopilot" when they still can't reliably avoid parked cars.

Ok the freeway, my Model 3’s autopilot is amazing. It’s not 100% on surface streets, but it’s impressive given all the edge cases and the scale of the problem.

They get around that by saying that autopilot is not meant to pilot the car automatically.

It’s clearly marked as in development. What’s impressive to me is just how good the 80% solution they have is, even without the additional FSD package.

It's an order of magnitude better at specific tasks compared to Nvidia.

And this isn't that impressive in the slightest.

Is a custom chip really that big of an accomplishment for a company the size and scale of Tesla? I really don’t know. Maybe someone here has some insight into semiconductor design and fabrication and can provide some insight into just how hard it would be to build a chip similar to what Tesla has built.