Akhetonics gets new funding for contrarian bets on all-optical chips
Photonics – the field that supports light-based systems for data management – has a bright future, as the rise of AI requires better computing performance, but it is yet to be fully exploited in the new generation of chips. German startup Akhetonics hopes to change that. A seed funding round of 6 million euros (about $6.33 million) has been raised to fulfill this promise, TechCrunch can exclusively reveal.
While several companies work with photons in critical issues or point solutions that mix electronics and photography, Akhetonics – whose name is a portmanteau of Akhet, the Egyptian hieroglyph for “horizon”, and photonics – aims entirely to build general purpose chip. .
“General purpose” in this context means chips that can be used for all kinds of tasks and software applications. And since Akhetonics takes an all-optical approach that will be digital and compatible with existing workloads, unlike analog methods, it can be especially useful in areas that require high performance in real time, such as networking, avionics, and space.
Speed aside, energy efficiency is another way in which photonics can help – and one that ties in as much to geopolitics, as it does to chip sovereignty. “For us, the most interesting part is that we have a very diverse supply chain,” founder and CEO Michael Kissner told TechCrunch.
Potentially, Akhetonics can make its general-purpose chip anywhere, making it possible for companies to access high-performance computing that is available locally – if it works. That’s a big “if”; or more precisely, “when.”
Most observers agree that photonics will come to chips – but French VC firm Daphni, for example, recently said it would not invest in general-purpose chips just yet.
While Lightmatter, an imaging company that originally focused on chips, focused on networking to great success, it brought fast data transfer between CPUs and GPUs inside data centers.
Although it still seems far off to some, Matterwave Ventures, the VC firm that led the new round of Akhetonics, believes that the time is right for the perfect virtual technology to be used in general purpose computing. “For us, it seemed like there were enough things coming together to make this happen,” principal Silviu Apostu told TechCrunch.
This will still take time, but perhaps not as much as some might think; Akhetonics plans to deliver its first commercial product to customers by the middle of next year. Kissner is confident that it is guaranteed to happen thanks to his previous funding of deep tech VC Runa Capital in 2023. shown,” he said.
The key to the Akhetonics method – and what makes this possible, according to Apostu – is to rethink the architecture from the first principles.
“People think he needs billions [optical transistors],” Kissner added. “But with the right structure, you don’t.” For example, the company explained in a recent paper how it could do without the standard optimizations used by AMD, Intel and Nvidia in current generation chips.
This also makes the development process cheaper than conventional chips – hence the relatively small round size of the chip making business. Akhetonics said most of the seed funding will grow to 30 people as the team works to deliver prototypes to customers. “For us, it’s actually a lot of money,” Kissner suggested. “In our world, you can design a chip for €50,000.”
Low cost and local supply chain are the biggest difference compared to existing AI semiconductors; Kissner seems genuinely confused that the “billion-dollar Al industry” relies on chips made in troubled parts of the country. And another area of Akhetonics obviously concerns investors. “They really support our goal of building this European, almost democratic version of high-performance computing,” he said.
With all that said, questions remain about the commercial need for high-performance computing, and whether it might be better served by integrated photonics for certain use cases. But for Akhetonics, and competitors like LightSolver, all-optical chips are the best answer.
Source link