A few weeks after closing a $300 million round backed by Nvidia Corp., Etched Inc. today announced that it has raised an additional $700 million in funding.

Investment firm Jane Street led the deal. It was joined by Kleiner Perkins, Sequoia, Andreessen Horowitz, Tiger Global, Bain Capital Ventures and several others. The raise more than doubles Etched’s valuation to $21 billion.

Etched launched in 2021 with the goal of designing chips optimized for specific artificial intelligence models. The company later pivoted to making inference accelerators that support a variety of model architectures. According to Etched, its first prototype chip rolled off a Taiwan Semiconductor Manufacturing Co. production line earlier this year. The company signed up Jane Street, the lead investor in today’s round, as its first customer shortly thereafter.

The investment firm installed an Etched-powered rack in its data center last month. Separately, the chip startup built a 2-megawatt AI cluster in its San Jose, California headquarters. Prospective customers use the cluster to test its chips.

An AI accelerator’s performance is influenced by the amount of heat it generates. The lower its operating temperature when running demanding workloads, the more calculations it can perform per second. One of the most effective ways to lower a chip’s operating temperature is to decrease the voltage that runs through it.

Etched says that its chip’s math blocks, or math-optimized circuits, run at under half the voltage of most AI accelerators. As a result, the processor can provide “multiple times the FLOPs density” offered by rivals. FLOPs density is a measure of a processor’s performance.

Etched didn’t disclose how its voltage-optimized math blocks work. Large language models perform two main types of mathematical operations during inference: dot product calculations and matrix multiplications. The math blocks are presumably optimized for those tasks.

A dot product calculation turns two rows of numbers into a single number. Such operations power the transformer architecture’s attention mechanism. Matrix multiplications, meanwhile, involve mathematical structures that can be visualized as spreadsheet tables filled with numbers. They’re used by not only LLM attention mechanisms but also other model components.

Etched ships its silicon as part of custom-designed racks. One of the design’s main selling points is an internally developed interconnect that links together the chips in each appliance. According to the Wall Street Journal, the interconnect can complete some communications tasks that take rival chips 4,000 milliseconds in 700 milliseconds. A millisecond is one-1,000th of a second.

Etched’s appliances also contain other internally developed components. There are custom cold plates, which conduct the heat generated by the company’s chips to the coolant circulated through the host systems. A proprietary VRM, or voltage regulator module, optimizes the flow of electricity in the appliances.

“It took us three years to deliver our first rack from scratch,” said Etched co-founder and Chief Executive Officer Gavin Uberti. “Our next one will be much faster.”

Image: Etched