Photonics - Xscape: Building the eight-wavelength laser for AI data
Xscape Photonics develops custom photonic platform products designed for ultra-high bandwidth connections inside datacentres to power AI systems.
The company’s proprietary ChromX platform targets scaling of AI computing performance in what the company promises is an environmentally sustainable manner.
The technology is optimised for power, cost, scale and reliability. Essentially, the organisation is a semiconductor startup developing photonic solutions for next-generation AI datacentre fabrics.
Now driving to create greater momentum for this sector of the photonics market, the company has now being backed by Addition, a new investor, while Xscape also has with continued support from existing investors, including capitalisation-focused GPU company Nvidia, among others.
Arise FalconX
Xscape Photonics also announced the launch of FalconX: a fully redundant External Laser Small Form-factor Pluggable (ELSFP) device capable of emitting up to eight wavelengths or colours of light for ultra-fast, high-capacity and low-power optical data transmission.
An External Laser Small Form-factor Pluggable device is a compact optical transceiver module that houses a laser source externally (rather than internally within the module itself), which enables higher power, improved thermal management and longer-reach optical communications.
“Rapidly increasing bandwidth, power and cost demands of AI workloads have created a critical hardware bottleneck, forcing developers to use just a fraction of their GPUs’ capacity, thereby limiting the revolutionary potential of AI itself,” said Vivek Raghunathan, Xscape Photonics’ CEO and co-founder. “Xscape Photonics is accelerating the development of its multi-colour wavelength-division multiplexing (WDM) fabric solutions to escape these hardware limits and fundamentally reimagine how data moves through datacentre networks.”
Raghunathan further states that FalconX is the industry’s first Comb laser module in a pluggable form factor capable of generating eight wavelengths of light, powering high-speed data movement to allow the entire datacentre to function as one giant GPU.
Laser terms defined
By way of additional definitions, multi-colour wavelength-division multiplexing involves using multiple light wavelengths simultaneously to transmit parallel data streams. Further, a Comb laser module can be explained as a single laser generating multiple evenly spaced frequency channels simultaneously.
Xscape Photonics was founded in 2022 by Raghunathan and four photonics scientists from Columbia University: Alexander Gaeta (president), Yoshi Okawachi (vice president of R&D), Keren Bergman (board of advisors member) and Michal Lipson (board of advisors member).
The company says that, currently, AI inference performance is constrained by conventional datacentre networks that rely on outdated copper interconnects, significantly limiting the amount of data that can escape from each accelerator to neighbouring accelerators or memory within AI clusters.
FalconX helps solve this “escape bandwidth” bottleneck by employing Xscape Photonics’ proprietary CombX laser technology, which generates multiple wavelengths of light on a single silicon photonics chip and enables the integration of high-performance, multi-colour optical interconnects into AI datacentre networks.
Multi-terabits-per-second
With more than 1W of optical power from a single pluggable laser module, FalconX generates eight colours that can power multi-terabits-per-second of data bandwidth. AI clusters have grown more than tenfold in size over the last two years and the failure of a single laser module can significantly impact the network, stalling the workload and increasing the token cost. Hyperscalers demand 10 times fewer failures compared to the incumbent laser level due to the growth of AI clusters. FalconX offers built-in redundancy and more reliable components to meet growing cluster demands.
FalconX is designed to comply with industry MSA standards for Scale-Up and Scale-Out fabric links and can be qualified in existing hyperscaler infrastructure.
