Semron, a German firm, is developing “3D-scaled” processors to run AI models locally on smartphones, earphones, VR headsets, and other mobile devices.
Semron’s chips, invented by Dresden University of Technology engineering grads Kai-Uwe Demasius and Aron Kirschen, calculate using electrical fields instead of electrical currents. Kirschen believes this increases energy efficiency and lowers manufacturing costs for semiconductors.
“Due to an expected shortage in AI compute resources, many companies with a business model that relies on such capabilities risk their existence—for example, large startups that train their own models,” Kirschen told Eltrys via email. “Despite our chips’ advanced AI capabilities, our technology will allow us to hit the price point of today’s chips for consumer electronics devices.”
Demasius and Kirschen submitted an initial patent for Semron’s chips in 2016, four years before they started the company. They use a “memcapacitor,” or capacitor with memory, to execute computations. Transistors, unlike capacitors, cannot store energy and operate as “on/off” switches to let or stop electric current in most computer chips.
Semron’s traditional semiconductor memcapacitors use charge shielding. Mescapacitors use a “shielding layer” to adjust the electric field between the top and bottom electrodes. The chip’s memory stores AI model “weights” and controls the shielding layer. As a model trains and processes data, weights work like knobs to manipulate and fine-tune its performance.
The electric field technique reduces chip-level electron mobility, energy, and heat use. Semron wants to use the electric field’s heat-reduction capabilities to stack hundreds of memcapacitors on a chip, improving computing capability.
“We use this property as an enabler to deploy several hundred times the compute resources on a fixed silicon area,” Kirschen said. “Think of it like hundreds of chips in one package.”
A 2021 Nature Electronics study by Semron and the Max Planck Institute of Microstructure Physics taught a computer vision model with an energy efficiency of about 3,500 TOPS/W, 35 to 300 times greater than previous methods. TOPS/W is an imprecise statistic; however, memcapacitors can dramatically reduce energy usage when training AI models.
Semron is at the “pre-product” stage and has “negligible” income, according to Kirschen. Chip startups often struggle with mass production and client acquisition, but not always in that order.
Semron faces severe competition from custom chip businesses like Kneron, EnCharge, and Tenstorrent, which have garnered tens of millions of dollars in venture capital. Similar to Semron, EnCharge designs computer chips with capacitors instead of transistors and a distinct substrate architecture.
Semron, which has 11 employees and plans to hire 25 by the end of the year, has raised money from Join Capital, SquareOne, OTB Ventures, and Onsight Ventures. Currently, the firm has raised 10 million euros (~$10.81 million).
Said SquareOne partner Georg Stockinger via email:
The 21st century’s “oil” will be computing resources. With infrastructure-hungry, huge language models overtaking the world and Moore’s law hitting physics’ limits, a colossal computational bottleneck will influence the future. Companies and nations will lose productivity and competitiveness without enough computer infrastructure. Semron’s breakthrough AI model-specific microprocessor will help solve this challenge. It violates the transistor-based computer paradigm and cuts processing costs and energy usage by at least 20x.”
