In the high-stakes world of artificial intelligence, advancements in hardware often go hand in hand with breakthroughs in software. Most AI experts concur that the next significant stride in AI could be largely contingent on the creation of supercomputers of unparalleled scale. Lightmatter, an emerging startup, believes it may hold the key to this ambitious leap forward with a groundbreaking technology that enables chips to communicate using light, potentially reshaping the landscape of hyperscale computing.
At a recent Sequoia venture capital event, Lightmatter’s CEO introduced their ambitious project called “Passage,” a transformative approach aimed at directly connecting GPUs—the workhorses of AI training—through optical links. Traditional computers and data centers typically transfer data via electrical signals. However, this process can introduce a significant communications bottleneck, particularly when signals are converted between optical and electrical forms for enhanced bandwidth.
Lightmatter’s Passage technology promises to circumvent these issues by leveraging optical or photonic interconnects fabricated in silicon. These interconnects facilitate direct interfacing with the transistors on a silicon chip, like a GPU, surging the data transfer bandwidth by a factor of 100 compared to current technologies.
For perspective, GPT-4, OpenAI’s flagship AI algorithm which powers ChatGPT, reportedly ran on a setup exceeding 20,000 GPUs. With Passage, which is expected to be operational by 2026, the scalability could leap to over a million GPUs working in tandem during an AI training session.
The potential of such technology has not escaped notice. Sam Altman, CEO of OpenAI, attended the Sequoia event. Altman has been vocal about the necessity of building larger and faster data centers to drive the progression of artificial intelligence forward. In fact, he has reportedly sought massive funding upwards of $7 trillion to invest in chips for AI, pointing to an engagement with Microsoft for a potential $100 billion data center endeavor, codenamed “Stargate.”
One major concern about present data centers is their energy consumption, particularly as the interconnectivity of chips on a grand scale demands a staggering amount of electricity. Optical communication, as proposed by Lightmatter, could present a viable solution to this energy bottleneck.
GlobalFoundries, a chip manufacturer for various prominent clients like AMD and General Motors, has acknowledged the potential of Lightmatter’s technology through a partnership. The startup claims to be working with the leading semiconductor companies and major cloud service providers, which could spell a major shift in data center design and operation.
Should Lightmatter succeed—or any company with a similar vision in reinventing the connectivity within AI mega-projects—the result could clear a significant hurdle in the development of sophisticated algorithms. The grandeur goal of achieving artificial general intelligence (AGI), or the capacity for AI to match or surpass biological intelligence in all aspects, could be considerably closer with Passage.
The use of optics in AI data centers could simplify the labyrinth of connections currently involving numerous switches and layers. In an optically connected center, each GPU would maintain a direct high-speed link to every other chip, streamlining communication and potentially revolutionizing large-scale AI training.
While Lightmatter captures the imagination with its Passage technology, veteran players in the industry like Nvidia continue to push the envelope. At a recent conference, Nvidia unveiled Blackwell, a powerful GPU for AI training, and announced that it would be sold in a “superchip” format. The innovation from Nvidia highlights continued efforts to enhance AI-supercomputer capabilities, indicating that technologies like Passage could become even more vital to future developments.
In summary, AI is set to grow by leaps and bounds, led by visionaries like Lightmatter who dare to think beyond the status quo. By harnessing light for communication between chips, the horizon of artificial intelligence could expand into realms once deemed the purview of science fiction, potentially ushering us into a new era defined by the realization of AGI.
