Table of Contents
Amid the 1.6T demand surge, interconnects are emerging as first-class citizens of the infrastructure
The race to scale artificial intelligence is increasingly forcing data center operators to find faster means to move data across sprawling, high-performance networks. The need has brought into being the 1.6 Terabit class of connectivity, which, if you’re at OFC this year, you may have heard repeated ad infinitum. With AI training and inference workloads outgrowing single servers, and expanding to rows of racks and even multiple data centers, the challenge of keeping the network working reliably is shifting toward interconnects.
1.6T test and validation
“For us, one of the biggest trends is the 1.6T,” said Manodipto Ghose, product manager at Keysight, during an interview with RCR at OFC where Keysight showcased its interconnect error-performance validation portfolio. “We demoed our first 1.6T last year at OFC, but this year we’re seeing a lot of traction.”
The traction is closely tied to the maturation of 224G SerDes which makes the foundational block for 1.6T components. “224G SerDes is going into mass volume production,” said Charles Seifert, Sr. manager of product management. “As that happens, the number of optical transceivers and active copper cables has proliferated, and there’s a desire to build those in volume.”
At its core, 1.6T Ethernet is about enabling faster data movement, a key requirement of distributed AI workloads. “1.6T is the base,” Ghose said. “The Ethernet connectivity, the scale-up networking—everything comes into this 1.6T Ethernet sphere.”
In scale-up and scale-out environments where thousands of GPUs operate as a single system, latency, packet loss, and synchronization are existential challenges. “Missing one packet in that collective is very expensive today,” Ghose noted.
This puts pressure on the interconnects — precisely their reach, power efficiency, and interoperability. Technologies like active copper cables (ACC) are gaining prominence as they extend connectivity beyond servers.
“Active copper allows us to send 200G signals over longer distances—two, three, even five meters, which is important for AI infrastructure racks,” Seifert explained.
The first-order priority is now validating these interconnects, which was once strictly a lower-layer conern. As Seifert put it, “There’s a higher emphasis than ever before on interconnect reliability and error performance.”
He explained, “If a cable has error performance problems and the link flaps, that part of the network goes down. That can cost up to $500,000 an hour.”
Keysight’s new 1.6T validation portfolio supports a variety of interconnect types critical for both network architectures, including 1.6T-capable passive copper Direct Attach Cables (DAC), ACC, Low Power Optics (LPO) and Linear Receive Optics (LRO).
Ultra Ethernet
Seifert also talked about the importance of Ultra Ethernet for AI-scale networking. Traditional Ethernet, he said, is build for “best-effort” delivery, where if packets don’t get through the first time, they are retransmitted. That is now being re-engineered for AI workloads.
“Ultra Ethernet is attempting to bring lossless Ethernet and guaranteed delivery,” Seifert said, referencing efforts by the Ultra Ethernet Consortium (UEC) of which Keysight is an early principal member.
The effort is aimed at closing the gap with InfiniBand, another high-speed open-standard interconnect technology used in data centers. “There is a silent war going on between Ethernet and InfiniBand,” he said. “But we’re already one speed ahead at 1.6T.”
Interesting, Ghose noted that some of the traditional network testing methods are making their way back, albeit in a different form.
“Some of those high port-count tests are coming back,” he said, refering to equipment with 64, 128, and 256 ports “We are taking a lot of learning from the traditional methods, but we are throwing in new pieces at it like RoCEv2 and Ultra Ethernet,” creates what is a hybrid testing approach, much like the AI network fabric itself.
One thing often discussed in conversations today is the giddy pace of technology interrations spurred by AI. This is palpable on the supplier side too.
“We’ve not even finished the IEEE standard on 224G, and we’ve started 448G,” Seifert said, adding that beyond 1.6T, companies’ roadmaps already include plans for 3.2T and 6.4T.
“They’re all overlapping the same two to three year timeframe..We’re shipping products that are not done to customers who are not done, for technologies that are not done,” he said.
