(Credit: Rob Pegoraro)
As people take an increasingly dim view of data centers, rewiring those computing facilities with high-speed, low-latency photonics links might let them have their cloud and AI services without as much air pollution and escalating electric bills.
The basic idea: By replacing traditional cables, switches, and other networking gear with optical equivalents that encode information as pulses of light, photonics could deliver vast advances in speed, latency, and power efficiency. But first, the hyperscalers who have been rushing to keep up with demand for data-center computing power—at the cost of corporate sustainability goals and sometimes environmental regulations—will have to take notice.
A Glowing Sales Pitch
Executives with such photonics efforts as NTT Data's Innovative Optical and Wireless Network (IOWN) project sketch out a new world of data centers in which operators can place them next to the cleanest and most abundant power instead of optimizing locations for bandwidth.
“You're not really interested if your GPUs are on the rack just below or 100 kilometers away,” Marco Provolo, head of the IOWN Global Promotion Office at NTT Data, said in an interview at the Japanese telco’s Upgrade conference in San Jose, California.
That reflects how NTT’s “all-photonics-network” eliminates traditionally wired switches and routers at network nodes: “You don't do any electrical-photonic conversion, which is actually the bottleneck," he added.
At Upgrade, NTT researchers and execs recited such longstanding IOWN goals as a 125x increase in capacity compared with current networks, reducing latency to 1/200th of today’s level, and improving energy efficiency by 100x.
They also touted such recent projects as a test of distributed AI training between two data centers about 22 miles apart in Japan, which saw the work take only 0.5% longer than on-premises training when split over an all-photonics link, versus 4.66 times longer over a conventional internet link.
Provolo cited other, longer-range tests of IOWN communication from Tokyo to Hiroshima, about 400 miles, and from Tokyo to Taipei, about 1,320 miles. “Taiwan is lacking GPUs, so we are offering media production remotely using GPUs in Japan,” he said.
Is Now a Better Time?
People have been talking a big game for photonics since the tail end of the previous century. In 1995, PCMag amplified a prediction of photonics in 2005’s consumer desktop PCs, and 10 years ago, Intel touted photonics upgrades for data-center infrastructure.
But beyond advances in engineering and design, the escalating demand for AI computing resources has radically upended the market, prompting people to consider ideas as out-there as space-based data centers. (NTT Research President and CEO Kazu Gomi’s take on that concept: “You can interpret this discussion as evidence that everybody is kind of desperate.”)
Will Townsend, chief analyst at LoneStar Advisory & Research, agreed in an email that skyrocketing demand for “more sustainable and higher performance ways to move data around” for AI services was making photonics more relevant.
The highest-profile company in AI processing, Nvidia, is now doling out some of its growing pile of cash to invest in photonics: In March, the firm announced a pair of $2 billion investments in two startups developing photonics for AI infrastructure, Coherent and Lumentum.
Space itself has become a proving ground for photonic technology, in the form of the laser links that SpaceX developed for its Starlink satellite constellation.
And on the ground, where atmospheric interference can make sending data via optical beams trickier, Taara, a spinout of X, Google's Moonshot Factory, has developed a compact silicon-photonics chipset that can support up to 25Gbps data transfers as far as 6.2 miles.
“We are taking the same kind of semiconductor technology and using optics on top of it,” CEO Mahesh Krishnaswamy said in an interview at Web Summit Vancouver.
Taara began as a successor to Google’s Loon project to deliver broadband from stratospheric balloons. But unlike that effort, which Google wound down in 2021, it’s now selling to a variety of customers, data center operators among them. Krishnaswamy’s sales pitch: “It is for real.”
Shipping When?
NTT’s long-running campaign for IOWN has, in the past, suffered from a lack of specifics about things like costs and ship dates. At Upgrade, NTT emphasized how, after showing that all-photonic links between data centers work (but without disclosing how many of its own are now IOWN-linked), it will soon ship the next phase of IOWN, optical links between boards in data centers.
“We are going to release a commercial product in Japan by the end of this year,” Gomi said, with subsequent phases to link processors and memory on a board with optics and link individual chip components.
But up-front costs remain somewhat vague in NTT’s public descriptions. Asked about them, Provolo said, “The expectation is that in a couple of years, the cost of this solution will decrease.”
He added that NTT plans to sell photonic interconnects between data centers as a service, citing financial-sector firms that need the lowest possible latency in their networks as a likely customer group. Provolo also said NTT continues to work toward ensuring IOWN does not leave potential customers seeing it as a vehicle for vendor lock-in. “The idea of IOWN is to create open standards.”
An industry consortium called the IOWN Global Forum is part of that idea; NTT says it’s gained 29 new members since a year ago, with Broadcom the highest-profile firm among them.
Townsend pointed to that group’s extensive membership as a strength for IOWN but also observed that two firms he called “front runners” in photonic interconnects for data centers, Lightmatter and Celestial AI (now a division of Marvell Technology), were not among those ranks.
But the analyst wasn’t willing to predict large-scale uptake of photonics in data centers before the end of the decade. That’s not because data-center operators, many of whom have publicly committed to paying for both the power they use and the additional infrastructure required to deliver it, couldn’t use its features, but because they want to see others prove it as a commercial reality first.
“What IOWN promises is audacious—dramatically lowering power consumption while increasing the speed of connectivity in equally dramatic fashion,” he wrote. “Until it gets to production, the hyperscalers will continue to focus on traditional data center deployments.”
Editors' Note: We updated this post to correct details about NTT and Taara photonics performance, Provolo's title and Taara's current relationship with Google.
Disclosure: NTT covered travel costs for journalists and analysts invited to this event; as a panel moderator at Web Summit Vancouver, that event's organizers paid for my hotel and airfare.