Salience Labs has begun shipping a 32-port optical circuit switch with a sub-10-nanosecond port-to-port hop.

The OC-32M device undercuts Broadcom's Tomahawk Ultra Ethernet ASIC by 25 times on switch latency. Released in March, the switch targets AI scale-up fabrics capped at 72 GPUs by Nvidia and AMD today.

Why does latency govern scale-up memory?

Standard Ethernet hops span 250 to 650 nanoseconds. MEMS-based optical circuit switches need milliseconds to reconfigure a port, leaving silicon photonics as the only viable path to memory-fabric-grade switching at scale inside an AI datacenter.

"We chose to develop OCS because our view is that there was quite a large number of players developing solutions on the CPO, NPO, XPO front, trying to solve the question of how you get data off of the chip and onto the optical fiber," CEO Vaysh Kewada told The Next Platform.

What sits inside the OC-32M?

The device stacks two chips on a back-to-back PCB card. One chip is a fully integrated silicon photonics OCS. The second is an amplification and signal conditioning chip that offsets optical loss accumulated across the integrated die.

The OCS uses 100 Gb/s native line rate with PAM4 modulation. Two bits per signal lift bandwidth to 200 Gb/s per lane. The architecture scales to 64, 128, and 256 ports, the company said.

Tower Semiconductor fabricates the PH18DA III-V lasers and TPS45PH low-loss silicon nitride waveguides for the device. Salience Labs has not disclosed its specific switching mechanism, though its founding research base and silicon nitride waveguide choice point toward a phase change material approach rooted in work by Oxford professor Harish Bhaskaran and Wolfram Pernice at Münster.

Who competes today?

Lumentum sells MEMS-based switches and has partnered with Nvidia on optics. Coherent uses liquid crystal on silicon. Google has wired "Palomar" MEMS devices into the "Apollo" OCS for four TPU generations since 2015, lashing up to 9,216 TPUs into a coherent memory cluster.

"Our view is OCS was a market ripe for disruption, because the OCSes that are available today on the market are based off a technology that fundamentally is twenty years old if not older," Kewada said.

Salience Labs claims 8X lower energy per port than a fast Ethernet switch. Reconfiguration runs more than 3X faster than MEMS devices on the company's own benchmark charts.

What is the scale-up ceiling now?

Nvidia and AMD both top out at 72 GPUs for coherent memory domains today. Salience Labs says prospective customers want scale-up beyond a single rack, alternative memory fabrics for rackscale machines, or OCS for the first network tier gluing machines together during GenAI decode phases.

The company modeled a 576-GPU cluster running Nvidia's Megatron 2T model. Switching at the scale-up layer lifted token-per-second-per-user throughput by 80%, without forcing a move to a dual-system architecture.

The startup can package up to eight OC-32M modules into one 1U chassis. Pricing and customer names remain undisclosed, with 64- and 128-port variants now sampling and 256 ports on the public roadmap, leaving hyperscale AI operators to test the silicon photonics switch against incumbent Ethernet and MEMS OCS designs through 2026.