Mixx Technologies shows its high-density SxC connector

Mixx Technologies is showing an expanded-beam connector that speeds up installing a rack’s optical connections at ECOC.

• The 64-fibre expanded-beam SxC connector addresses growing fibre counts in racks for scale-up networking.
• The connector supports silicon-coupled optics, front-panel fibre, and optical backplanes in a rack.
• Mixx’s connector uses wafer-scale manufacturing, and its glass-based design works in high-temperature environments and with high-power lasers..

Hyperscalers have built robots to seat servers into racks. Not because the server is too heavy for the technician, rather mating several thousand fibre connections at once requires more force than an individual can apply.

“The sheer force of that insertion process is too large for a human to do with today’s fibre connectors,” says Rebecca Schaevitz, Mixx’s co-founder and chief product officer.

It is a problem that customers brought to Mixx Technologies’ attention and explains why the start-up’s first product is an optical connector, given its focus is optical engine-based systems.

An expanded-beam connector does not need the fibre end faces pressed into physical contact, says Mixx. The force needed to mate its connector is a small fraction of that required using existing connectors.

At the Semicon Taiwan show held earlier this month, visitors were able to push and release a filled bulkhead by hand. A fully populated bulkhead has roughly 1,000 fibre mates.

It was important for Mixx to launch the SxC fibre connector at Semicon Taiwan. “This show is pretty much the Mecca of semiconductors, and all the discussion is about co-packaged optics and how to integrate it into the AI infrastructure,” says Vivek Raghuraman, CEO of Mixx. “The product we have released is critical here because the rack has become the new GPU, the new endpoint.”

Mixx also engaged with original design manufacturers (ODMs) and partners that make rack platforms, as well as chip players building co-packaged optics (CPO).

Mixx demonstrated how a board with 1,000 fibres could be inserted and mated with a single click, what it calls ‘scaling optics at the rack level’.

“A lot of people don’t necessarily understand connectors,” says Schaevitz. “What landed [with them] was the backplane implication — sliding a server into position and having every lane come up with no further action.” She adds that with the connections, it is not just an issue of achieving higher density; it is also the time it takes to make all the connections that matter, connections that can span multiple racks.

AI inferencing and training needs

Much is happening within compute racks and how the chips inside are connected, whether they are AI accelerators, large Ethernet switch chips, or memory.

Designers are focussed on integrating optics within the rack to enable one or more racks to be linked, thereby expanding the scale-up network. Here, optics is being introduced to complement or replace copper using co-packaged or near-packaged optics (NPO). This is to ensure AI workloads continue to scale, whether for model training or inference.

Schaevitz draws a sharp line between AI training and inference systems.

“In AI training, you’re trying to do as much as possible in a certain amount of time,” she says. “So, whatever you can do in six months, that is what that model will be.”

In AI inference, the goal is to serve many simultaneous users as quickly as possible at the lowest cost. “The industry as a whole has boiled that down into dollars per token per second,” she says. What matters here are system cost, system utilisation, and latency, the last two of which are networking issues.

Mixture-of-experts and agentic workloads also require memory. Schaevitz describes clusters specialising by task, much as regions of a brain do, each holding short- and long-term memory and each needing to talk to the others to assemble an answer. Every one of those relationships is a connection — to a GPU, TPU or xPU, or to the memory itself.

Schaevitz points out that these AI inferencing systems are going beyond a single rack, and that using existing 16-fibre connectors is challenging due to the sheer number of connections required.  A next-generation 204.8-terabit Ethernet switch with 200-gigabit lanes needs 2,048 fibres. On a backplane, that is 128, 16-fibre connectors. At 400 terabits, requirement doubles. At the front panel, 128 such connectors take up about half of a one rack unit (1RU) faceplate before any room is found for the lasers a co-packaged design also needs, so the box has to grow beyond 1RU or use both front and back, which brings back the backplane force problem.

To this end, Mixx’s SxC connector uses 64 fibres, arranged in a 2D array. Mixx claims that, using its connector, a next-generation switch will fit in a one rack unit (1RU) box. Otherwise,  a four-rack-unit (4RU) platform would be required using existing MMC (Miniature Multi-Fiber) or other connectors, says Raghuraman.

In turn, a spine switching layer built using such switches needs four times the rack space if each switch occupies 4RU rather than 1RU. A spine layer that would require several racks using existing connectors can now fit into one rack using the SxC connector, says Mixx.

With the advent of near-packaged and co-packaged optics, certain end users will adopt optical backplane-based racks, a direct analogue of the copper spine found in AI rack systems such as Nvidia’s NVL72. Once optics moves inside the box, the optical spine will deliver equivalent point-to-point connectivity. The optical version can also perform shuffling to handle fibre reordering.

The second benefit of Mixx’s connector is how it speeds up getting the platform up and running. At present, says Mixx, deploying a rack requires a technician to plug in the fibre connections one by one.

“Here you can eliminate that time needed to connect the fibres; you just install the rack and boom, you improve deployment time significantly,” says Raghuraman. “That is the value of the backplane, but there is no solution that has enabled it to date.”

Expanded Beam Optics (EBO) Multi-source Agreement (MSA)

Mixx Technologies is also part of the Expanded Beam Optics (EBO) Multi-source Agreement (MSA), launched in May, whose members include 3M, Arista, Cisco, Meta, Molex and HPE.

The MSA, which now has over 50 members, aims to create open, standard rules for optical connections in AI data centres. Using an expanded beam simplifies fibre connections and helps mitigate issues such as dust and contaminants making connections more reliable.

Raghuraman says Mixx’s connector focus is currently less about interoperable form factors and more about its expanded-beam solution to tackle issues such as dust, mating and demating, and achieving low insertion loss and higher power handling.

Mixx notes the SxC uses glass and not plastic or polymer parts, such that key elements of the connector – the lens, waveguides and ferrule – can be manufactured at wafer scale. And being glass-based, the connector can handle high temperatures and is designed to handle optical power of up to 400 milliwatts.

“Use cases are evolving where the laser goes remote and not as an ELSFP,” says Raghuraman. “In those scenarios, you are going to have very high power going through these networks. Industry-standard connectors built from organic materials are limited to 50 milliwatts, says Mixx.

View source version on gazettabyte: https://gazettabyte.com/mixx-technologies-shows-its-high-density-sxc-connector/