A new startup based in Germany, Semron, is pioneering "3D-scaled" chips designed to enable local AI processing on mobile devices such as smartphones, earbuds, and VR headsets. Co-founded by Kai-Uwe Demasius and Aron Kirschen, both engineering graduates from the Dresden University of Technology, Semron's innovative chips utilize electrical fields for computations rather than electrical currents, which are typically used in conventional processors. This groundbreaking approach significantly enhances energy efficiency while keeping manufacturing costs low, according to Kirschen.
"Given the anticipated shortage of AI compute resources, many companies that depend on these capabilities, especially large startups training their own models, face existential risks," Kirschen explained in an email interview. "Our technology's distinctive features allow us to match the pricing of current consumer electronics chips, even though our chips can manage advanced AI operations that others cannot."
Demasius and Kirschen initially filed a patent for Semron’s technology in 2016, four years prior to launching the company. Their chips incorporate a unique component termed a "memcapacitor," which functions as a capacitor with memory. Unlike traditional transistors that merely act as switches and cannot store energy, Semron’s memcapacitors utilize semiconductor materials to support advanced calculations. The memcapacitors manipulate an electric field between their electrodes via a "shielding layer," the properties of which are determined by the chip's memory. This memory stores the varying "weights" used in AI models, which effectively tune the model's performance during training and data processing.
The use of electric fields reduces electron movement within the chip, resulting in lower energy consumption and minimized heat production. Semron aims to capitalize on these heat-reducing properties, potentially allowing for hundreds of memcapacitor layers to be integrated on a single chip, thereby vastly improving computational capacity.
“We leverage this unique characteristic to deploy hundreds of times the computational resources in a fixed silicon area," Kirschen noted. "Imagine having hundreds of chips encapsulated in one package."
In a 2021 study published in Nature Electronics, researchers from Semron and the Max Planck Institute of Microstructure Physics demonstrated successful training of a computer vision model, achieving energy efficiencies exceeding 3,500 TOPS/W—35 to 300 times more efficient than existing techniques. Although TOPS/W is a fairly abstract measurement, the key takeaway is that memcapacitors significantly reduce energy consumption while training AI models.
Currently, Semron is in the "pre-product" phase and reports minimal revenue. One of the biggest challenges for chip startups like Semron is scaling up mass manufacturing and building a robust customer base. Compounding the challenge is the stiff competition from custom chip companies such as Kneron, EnCharge, and Tenstorrent, which have collectively secured tens of millions in venture capital. While EnCharge shares a similar focus on capacitor-based chips, it employs a different substrate architecture.
Despite these hurdles, Semron has attracted investment from notable firms, including Join Capital, SquareOne, OTB Ventures, and Onsight Ventures. To date, the startup has raised approximately 10 million euros (~$10.81 million).
SquareOne partner Georg Stockinger commented, “Computing resources are destined to become the ‘oil’ of the 21st century. As infrastructure-intensive large language models proliferate and Moore’s law approaches its physical limits, a severe bottleneck in computing resources will manifest in the coming years. Insufficient access to computing infrastructure will hinder productivity and competitiveness for companies and entire nations alike. Semron is poised to play a pivotal role in addressing this issue by providing a revolutionary chip specifically designed for AI model computing, breaking away from the traditional transistor-based paradigm while achieving at least 20 times the cost and energy efficiency for each computational task.”
Ranking
