Co-Founder & CEO at Zettascale
Energy efficient chips for AI
Zettascale builds XPUs, reconfigurable chips that optimize their dataflow for each AI model instead of running a fixed function. Nearly all the energy in AI goes into moving data, so the architecture is built to keep it close. Elias has been an autodidact since he was nine. He left a computer science and computer engineering degree at Chalmers to build the company, first named Exa Laboratories. Its YC launch put the XPU at up to 27.6 times more efficient than an H100 in early simulation. Grasshopper, the first XPU, is prototyped on an FPGA, and Monolith links XPUs so a cluster behaves as one chip. Zettascale (YC S24) is backed by Soma Capital and Olive Tree Capital.
Elias Almqvist is the co-founder and CEO of Zettascale, which builds energy-efficient chips for AI. He has been an autodidact since the age of nine and left a computer science and computer architecture engineering degree to start the company.
Zettascale's chip. Unlike a fixed-function GPU or TPU, an XPU is highly programmable, so it can reconfigure its dataflow for each model. That is where the efficiency comes from: dense math, sparse models and the precisions they actually use, in one architecture.
The company's YC launch put the XPU at up to 27.6 times more efficient than an H100 GPU, based on early simulation rather than shipped silicon.
Grasshopper is the first XPU, prototyped on an FPGA. Monolith is a cluster of XPUs that behaves as a single chip, aimed at agents, experience generation and training on the same silicon.
Because the energy cost of training and running large models is becoming the constraint on data centers. Zettascale's pitch is that a chip which moves less data saves data centers a large share of their energy and cooling bill.
Yes. Zettascale Computing Corporation launched as Exa Laboratories in Y Combinator's S24 batch and later changed its name.
