Insufficient onboard processing power? Tesla’s HW3 computer is about 70Tflops, ~0.1% of estimated 100Pops human brain. Approximately equivalent to a mouse brain. Social and predator mammals that have to model and predict conspecific and prey behaviors have brains that generally start at about 2% human for cats and up to 10% for wolves.
I posit that driving adequately requires modelling, interpreting and anticipating other road users behaviors to deal with a substantial number of problematic and dangerous situations; like noticing erratic/non-rule compliant behavior in pedestrians and other road users and adjusting response, or negotiating rule-breaking at intersections or obstructions with other road users, and that is orders of magnitude harder than strict physics and road rule adherence. This is the inevitable edge-case issue that there are simply too many odd possibilities to train for and greater innate comprehension of human behaviors is needed.
Perhaps phone-a-smarter-AI-friend escalation of the outside-of-context problems encountered can deal with much of this, but I think cars need a lot more on board smarts to do a satisfactory job. A cat to small-dog level H100 @4Pflops FP8 might eventually be up to the task.
Elon’s recent announcements about end-to-end neural nets processing for V12.xx FSD release being the savior and solution appears to not have panned out—there’s been months of silence after initial announcement. That may have been their last toss of the die for HW3 hardware. Their next iteration HW4 chip being fitted to new cars is apparently ~8x faster at 600Tflops.[edit, the source I found appears unreliable, specs not yet public] They really wouldn’t want to do that if they didn’t think it was necessary.
Insufficient onboard processing power? Tesla’s HW3 computer is about 70Tflops, ~0.1% of estimated 100Pops human brain. Approximately equivalent to a mouse brain. Social and predator mammals that have to model and predict conspecific and prey behaviors have brains that generally start at about 2% human for cats and up to 10% for wolves.
I posit that driving adequately requires modelling, interpreting and anticipating other road users behaviors to deal with a substantial number of problematic and dangerous situations; like noticing erratic/non-rule compliant behavior in pedestrians and other road users and adjusting response, or negotiating rule-breaking at intersections or obstructions with other road users, and that is orders of magnitude harder than strict physics and road rule adherence. This is the inevitable edge-case issue that there are simply too many odd possibilities to train for and greater innate comprehension of human behaviors is needed.
Perhaps phone-a-smarter-AI-friend escalation of the outside-of-context problems encountered can deal with much of this, but I think cars need a lot more on board smarts to do a satisfactory job. A cat to small-dog level H100 @4Pflops FP8 might eventually be up to the task.
Elon’s recent announcements about end-to-end neural nets processing for V12.xx FSD release being the savior and solution appears to not have panned out—there’s been months of silence after initial announcement. That may have been their last toss of the die for HW3 hardware.
Their next iteration HW4 chip being fitted to new cars is apparently ~8x faster at 600Tflops.[edit, the source I found appears unreliable, specs not yet public] They really wouldn’t want to do that if they didn’t think it was necessary.Helpful comment that gives lots to think about. Thanks!