While the transistor-based silicon chip has greatly increased computing power in recent decades as transistors have grown to a width of many atoms, shrinking them further is a challenge. Not only is it difficult to make something so small, but when you get smaller signals it can leak in between. Thus, Moore’s Law, according to which the density of transistors on a chip must double every two years and costs decrease, is slowing down, prompting the industry to search for new solutions to meet the growing needs in AI computing.
According to data firm PitchBook, silicon photonics startups raised more than $750 million last year, twice as much as in 2020. And in 2016, that was about $18 million.
“Artificial intelligence is growing like crazy and is taking over large portions of data centers,” Charles Wishpard, CEO of Ayar Labs, said in an interview with Reuters. “The challenge of data traffic and the power consumption of this data traffic is a big, big problem.”
The problem stems from the fact that many large machine learning algorithms may use hundreds or thousands of chips for computation, and there is a bottleneck in the speed of data transfer between chips or servers using existing electrical methods. Light has been used for decades to transmit data over fiber-optic cables, including subsea cables, but getting it to chip level has been difficult because the devices used to create or control light are not as easy to shrink as transistors.
Brendan Burke, Senior Analyst for Emerging Technologies at PitchBook, expects silicon photonics to become mainstream devices in data centers by 2025 and estimates the market to reach $3 billion by that time, comparable in size to the AI graphics chip market in 2020.
In addition to connecting transistor chips, startups using silicon photonics to build quantum computers, supercomputers and chips for self-driving vehicles are also raising big money.
PsiQuantum has raised about $665 million so far, although the promise of quantum computers to change the world is still a long way off. Lightmatter, which makes processors using light to speed up AI workloads in the data center, has raised $113 million and will release its chips later this year and test them with customers soon after. Luminous Computing, a startup that builds an artificial intelligence supercomputer using photonic silicon, backed by Bill Gates, has raised a total of $115 million.
Startups aren’t the only ones pushing this technology forward. Semiconductor makers are also preparing to use their silicon wafer manufacturing technology for photonics.
The collaboration with PsiQuantum, Ayar and Lightmatter has created a silicon photon manufacturing platform that others can use, said Amir Fentoch, Head of Computing and Wired Infrastructure at GlobalFoundries. This platform was launched in March.
Peter Barrett, founder of venture capital firm Playground Global, which has invested in Ayar Labs and PsiQuantum, believes in the long-term prospects of silicon photonics to accelerate computing, but thinks it will be a long way to go.
“What the guys at Ayar Labs are doing really well… is they’ve solved the problem of linking data for traditional high-performance computing,” he said. “But it will take time before we have pure digital optical computing for non-quantum systems.”