https://www.optica-opn.org/home/articles/volume_35/september_2024/features/photonics_and_ai_industry_perspectives/?utm_content=305513552&utm_medium=social&utm_source=linkedin&hss_channel=lcp-6627049

Hannah Lanford, Rachel Sender and Stewart Wills 

OPN talked with six companies about some of the opportunities AI could open for the optics and photonics industry—in silicon photonics and beyond.

The release of Open AI’s generative-AI chatbot, ChatGPT, in November 2022 was the technology shot heard ‘round the world, opening what seemed to be new vistas of productivity and potential across the business landscape. And although much of the initial public hype has since been tempered, the continued rapid growth of AI-capable infrastructure in data centers, and of AI-enabled thinking in the product lab, is creating intriguing opportunities for the optics and photonics industry.

For a taste of how all of this is unfolding, OPN recently spoke with representatives from six companies looking at pieces of this marketplace. While the resulting snapshots tell only a small part of a very big story, they do hint at how generative AI’s data needs have catalyzed a ferment of activity in silicon photonics—and also at AI’s broader impacts on how optics firms think about their other product offerings.

Chiplet time

For silicon photonics, much of the action in serving these markets lies in solving AI’s “data bottleneck.” That’s shorthand for the increases in latency and energy consumption that occur as AI-enabling CPUs, GPUs and application-specific integrated circuits (ASICs) suck in data at an ever more furious pace from off-­package computational resources—memory modules, for example, or other ASICs—across copper interconnects.

Numerous companies are converging on a similar solution to that quandary: silicon photonic “chiplets” that push the conversion between the electrical and optical domains as close to the GPU as possible, allowing connectivity across distributed computing resources to stay encoded as fast, energy-efficient light in optical fiber. One company that has plied these waters for nine years is Silicon Valley–based Ayar Labs. Ayar was born in May 2015 out of work (later reported in Nature) from labs at the University of California, Berkeley, the University of Colorado and the Massachusetts Institute of Technology (MIT), USA, demonstrating what Ayar now calls “the world’s first processor to communicate using light.”

Terry Thorn—who joined the firm as vice president of commercial operations three years ago, after 24 years in various roles at the semiconductor giant Intel Corp.—told OPN that Ayar’s work focuses on problems in the “scale-up” domain inside compute systems. That’s the domain that involves shunting data between compute chips, be they a CPU and GPU, two GPUs or a GPU and a distributed memory host. (It contrasts with the “scale-out” domain involving data transmission out of the server rack to other servers, and typically the sweet spot for traditional pluggable transceivers.)

In the scale-up domain, when a myriad of traditional copper interconnects is replaced by optical links carried in fiber, “you benefit a lot, obviously, from reduced latency and reduced energy,” according to Thorn. “And the closer we can put our connection to a compute chip, the more we’re able to take advantage of that efficiency benefit.” It’s in the scale-up domain, he said, that Ayar sees “the greatest immediate benefit, and where we have the greatest urgency and interest from our customers.”

Ayar’s vehicle for addressing that interest is a chiplet with the slightly whimsical name TeraPHY (pronounced “terrify”). The chiplet is designed to be integrated in-package with the electronic GPU or CPU, allowing electrical data to be immediately encoded (via systems of microring resonators) into optical data and sent out to other, similarly equipped compute resources on ribbons of optical fiber.

At OFC 2023, Thorn said, Ayar showed a “very stable demonstration” of its chiplets running “in live silicon” in a 4-Tbps bidirectional connection (that is, 2-Tbps in each direction) that used an energy-stingy 5 picojoules (pJ) per bit. The high data throughput was enabled by wavelength-division multiplexing in eight fiber connections to the chiplet, each hosting eight wavelengths of light carrying data at 32 Gbps apiece. Eight months later, in November 2023, Ayar demonstrated a similar system packaged with Intel field-programmable gate arrays (FPGAs) “running live, for days on end,” according to Thorn.

The other key component in these demos, and in Ayar’s offering, is its off-chip light source, trade-named SuperNova, that plugs into the chiplet. While the version used in the OFC 2023 demonstration supported eight wavelengths, one year later, at OFC 2024, Ayar unveiled an upgraded 16-wavelength SuperNova that the company says can “drive 256 optical carriers for 16 Tbps of bidirectional bandwidth”—a level of zip that Thorn thinks the company will demonstrate publicly in the next several years. In addition to changing the number of wavelengths, Thorn said, Ayar can use “a number of other levers” to push throughput, including adding fiber ports to the chip and boosting the data modulation per wavelength from the current 32 Gbps.

(full article continues at link)

[Enlarge image][Celestial AI]

An active playground

Here’s a random sample of some of the many other firms attempting to carve out niches in the AI data space.

Celestial AI. The company’s offering is built around a “Photonic Fabric” that keeps data encoded as light not only to the edge of the GPU, but all the way to the “point of compute”—the precise point on the computer die where the number-crunching is taking place. The novelty of the concept has attracted substantial funding, as well as exploratory partnerships with what Celestial calls “some of the leading hyperscalers” as it works to build a “Photonic Fabric ecosystem.” [optica-opn.org/news/0524-fabric]

POET Technologies. The Toronto-based integrated-photonics firm is stressing its “Optical Interposer,” which POET calls “the first-ever wafer-level integration of electronics and photonics into a single device,” as a CMOS-compatible platform technology that can be used as a building block for co-packaged optics targeting AI applications at the transceiver and on-board levels.

Ranovus. Another Canada-headquartered firm, Ranovus, is specifically targeting optical interconnects in data centers through its Odin single-chip optical engine, which combines silicon photonics PICs with off-chip quantum dot lasers for AI/ML workloads. At OFC 2024, the firm announced that it was working with the Taiwan-based fabless semiconductor company MediaTek on “a 6.4-Tbps co-packaged optics solution” with “4-pJ/bit energy efficiency, including the laser.”

Quintessent. The California-based startup Quintessent has targeted laser sources as the “weakest link” for AI scalability, and hopes to strengthen that link through its own silicon photonics technology incorporating quantum dots and multiwavelength comb lasers. The company announced a US$11.5 million seed round earlier this year. [optica-opn.org/news/0624-quint]