One AI Chip Startup Says Its Chips Run on Under Half the Voltage

Gavin Urtz, a Harvard dropout building a new AI chip with his startup Etched, says the company is bringing forward the first generation of its low-voltage power system, which is "running at under half the voltage of any other AI chip." That lower voltage, he says, lets the company add much more math without overheating. The chips are not described as shipping yet.

The problem starts with wasted math. AI chips are rated in flops, or trillions of math operations per second, but what matters is how many of those operations you actually get when running a real chatbot or agent. That share has a name, model flops utilization, and on graphics chips it is often only 20 to 50 percent.

Heat is one major cause of the gap. When more transistors switch on and off, the chip draws more power and gets hotter, so it slows itself down to avoid overheating. Simply adding more math circuits does not help, because the chip will just throttle again.

That is why voltage matters so much. Under an old chip rule called Dennard scaling, power rises with the square of voltage. Double the voltage and power quadruples. Cut voltage in half and power falls to one quarter. Lower voltage is a way to pack in far more flops without overheating.

When Urtz and his cofounders asked experienced chip designers how to do that after dropping out, he says they were told it could not be done. He found that answer unconvincing, because other kinds of chips already do it. As Urtz put it, "Bitcoin miners run at under a quarter of the voltage of GPUs."

That example showed it was physically possible, Urtz says. His team then created what Urtz called "a new type of power delivery" for operating at far lower voltages, a system it calls low voltage inference. His prediction is that all future AI chips will be low voltage chips, packing in far more math on the same chip area and running them without overheating.