As the AI revolution continues to unfold, the capability to move data as quickly as possible into GPUs without driving up the power consumption–and thus the heat generation–has emerged as a major impediment to progress. This is driving the creation of a new generation of photonic-based interconnects and Ethernet switches that promise to minimize data bottlenecks in ultra-dense GPU clusters while keeping power and heat-generation to a minimum. Companies at the leading edge of photonics, including Broadcom, NLM Photonics, and Aloe Semiconductor, revealed this week they’re making progress.
The industry increasingly is looking to use Co-Packaged Optics (CPO) photonics technology as a replacement to traditional copper-based interconnects. CPOs integrate directly with processors and promise to deliver lots of data quickly, with lower latency, lower energy consumption, and lower heat generation than copper-based interconnects. Optics also has the advantage of suffering from less data loss as the runs get longer than traditional electrical signaling, which is becoming an issue as AI data centers get physically bigger.
At its GPU Tech Conference earlier this year, Nvidia revealed its first CPO offerings. When it ships in 2026, the Spectrum-X Photonics switch is expected to deliver 200 gigabits per lane of data connectivity on an Ethernet networks, while the liquid-cooled Quantum-X Photonics, which is expected by the end of 2025, will provide 800G connectivity for proprietary InfiniBand interconnects.
Another company at the forefront of CPO photonics tech is Broadcom, which integrates optics and the network ASIC on a single packaged substrate. Broadcom’s CPO can be packaged as an interconnect for HPC clusters running its XPU AI accelerators, or for use as an Ethernet switch for scale-out environments.
Broadcom’s shipped its first-generation CPO, dubbed Tomahawk 4-Humboldt, back in 2021, which offered a mix of optical and electronic channels, and delivered 100 Gigabits per second per physical optical link, or G/lane, and 25.6 Tbps total bandwidth. It followed that up in 2022 with its second-gen CPO, Tomahawk 5-Bailly, which was a full optical device running at 100G/lane, offering 51.2 Tbps total bandwidth.
Earlier this year, Broadcom announced its third-generation CPO offering based on Tomahawk-5, which was increased to 200G/lane. It’s now testing fourth-generation CPO based on its Tomahawk 6 processor, which delivers 400G/lane connectivity and which provides total bandwidth of 104.2 Tbps.
Broadcom is already shipping traditional Ethernet ASICS that move data over copper as opposed to strands of fiber optic glass. Its StrataXGS Ethernet switch is powered by its Tomahawk-6 chip and delivers 104.2 Tbps of total bandwidth. Tomahawk 7 and 8 chips are also in the works.
This week, Broadcom announced that Facebook’s parent company, Meta, has put its latest 400G CPO technology through rigorous testing in its high-temperature lab, paving the way for deliveries soon.
The test with Meta simulated running 400G port devices for more than 4 million hours and completed without experiencing a single “link flap,” which is network-speak for a brief connectivity disruption. That shows the tech will soon be ready for prime-time says Near Margalit, who is vice president and general manager of Broadcom’s Optical Systems Division.
“Achieving one million link flap-free hours is a strong validation of Broadcom’s commitment to quality and innovation,” Margalit stated in a press release. “This milestone shows that CPO is not just a research concept–it is production-proven and ready to scale.”
One of Broadcom’s industry partners, Micas Networks, can’t wait for the fourth-generation CPO. The company, co-packages the Broadcom switch along with eight 6.4-Tbps silicon photonics chiplets in package (SCIP) optical engines onto a single 4U rack offering that offers 128 ports of 400G FR4 fiber connectivity.
Another company pushing the photonics-in-AI-data-center ball forward NLM Photonics. The Washington State-based company this week announced the completion of testing for its silicon-organic hybrid (SOH) photonic integrated circuits (PICs), which are capable of delivering data at speeds of 1.6 and 3.2 terabits per second.
The tests clocked the SOH PIC delivering data at the rate of 224Gbit per lane, the company says, which demonstrates the viability of 200G performance in a commercial SOH package while paving the way toward 400G.
“The demand from AI data centers for high bandwidth and lower power, while not breaking the bank, is creating unprecedented challenges for the photonics industry,” NLM Photonics CEO Brad Booth stated in a press release. “Traditional photonic platforms are severely challenged to meet next-gen demands. Our 1.6T and 3.2T SOH PICs demonstrate the ability of NLM’s solutions to permit silicon photonic platforms to continue to scale to higher bandwidths while reducing a PIC’s footprint by at least 40%.”
New Jersey’s Aloe Semiconductor is also making progress in photonics. The company used the European Conference for Optical Communication (ECOC) Exhibition in Copenhagen, Denmark s the backdrop for its announcement that early tests show that its silicon photonics modulator achieved 180-Gbaud PAM4 (or 360 Gb/s).
The results “prove the strong likelihood of reaching 425 Gb/s without having to introduce new materials and processes into the foundries,” the company says.
“This modulator showcases the tenacity of silicon photonics,” Aloe’s CEO Christopher Doerr said in a press release. “With innovative design, silicon can reach higher and higher speeds while utilizing existing fabrication, allowing industry to continue to take advantage of the scaling and assembly abilities of silicon.”
Credit: Source link
