Taalas HC1: 17,000 tokens/sec on Llama 3.1 8B vs Nvidia H200’s 233 tokens/sec. 73x faster at one-tenth the power. Each chip runs ONE model, hardwired into the transistors.
Taalas HC1: 17,000 tokens/sec on Llama 3.1 8B vs Nvidia H200’s 233 tokens/sec. 73x faster at one-tenth the power. Each chip runs ONE model, hardwired into the transistors.
When the regular controller of the car - be it human, another AI, whatever - isn’t sending control signals, then the onboard controller knows that the car is uncontrolled. Of course it’s a “failure scenario”, I’m suggesting that this chip would be ideal for picking up when that sort of thing happens. The alternative is to just fall over.
I, too, am not sure what you’re arguing. I suggested that a low-power high-speed AI chip like this would be ideal for putting in robots, which have power constraints and aren’t always in reliable contact with outside controllers. That’s a very broad “niche” indeed. I don’t know what all this landmine stuff or probabilities of brake-slamming is all about or how it relates to what I suggested.