Passage
Passage is a 3D photonic interconnect platform from Lightmatter, designed to eliminate the bandwidth bottleneck in AI training and data center switching.
Publisher review
Passage is a 3D photonic interconnect platform from Lightmatter, designed to eliminate the bandwidth bottleneck in AI training and data center switching. It targets hyperscalers, AI infrastructure builders, and system architects who need to scale interconnect bandwidth beyond the limits of traditional electrical I/O and discrete optics. By embedding fiber-optic connections directly beneath processors, Passage enables massive bandwidth density across the entire die area rather than being confined to chip edges (shoreline). The platform is available in L-series (near-package, on-board, co-packaged) and M-series (interposer co-packaged) form factors, covering the spectrum from pluggable optics to full 3D integration. Lightmatter also offers the Guide light engine, based on Very Large Scale Photonics (VLSP) technology, which provides an integrated laser source to power Passage interconnects at hyperscale densities.
Passage's Edgeless I/O architecture places optical I/O across the full die area, allowing bandwidth density to scale with area (mm²) rather than perimeter. Specific products include: the Passage L200, supporting 32 to 64 Tbps aggregate bandwidth via co-packaged optics using 112G PAM4 signaling; the Passage L20, delivering 12.8 Tbps at 4x pluggable density; the Passage M1000 EVK, a rack-level 3D photonic interposer providing 114 Tbps across a 4,000 mm² footprint; the Passage EVK100, a bidirectional link delivering up to 3.2 Tbps per fiber at 1.9 pJ/bit; and the Passage EVK50, a 16-wavelength bidirectional link delivering 800 Gbps per fiber at 2.6 pJ/bit. The Guide light engine generates up to 51.2 Tbps per laser module for NPO and CPO applications, with a minimum of 100 mW per fiber across 16 multiplexed wavelengths, and enables bidirectional photonic links with precise 200 GHz wavelength spacing at +/- 20 GHz accuracy.
Passage competes directly with other photonic interconnect and co-packaged optics solutions from companies like Ayar Labs, Intel (with its silicon photonics), and Marvell (with its PAM4 DSPs and optics), though the provided sources list competitors such as Copy AI, Drift, Clari, Podium, and VERSE AI — which appear to be unrelated AI writing or conversational intelligence tools, not photonics companies. This suggests the extracted competitor list may be erroneous or drawn from a different context. In the photonic interconnect space, Passage's key differentiator is its 3D Edgeless I/O architecture and the VLSP-based Guide light engine, which reduces component counts and assembly complexity compared to discrete ELSFP modules. The first-generation Guide validation platform enables 100 Tbps of switch bandwidth in a compact 1RU chassis, a density that would require approximately 18 conventional ELSFP modules occupying 4RU of rack space.
Passage's primary trade-off is that it is early-stage technology — available only for early access partners, not yet in mass production — meaning deployment risk for hyperscalers who need proven, field-tested solutions today. The platform also requires significant infrastructure changes (new packaging, cooling, and fiber management) compared to pluggable optics, increasing initial deployment complexity. While the Guide light engine promises superior yields and reliability through foundry-style production, the actual field reliability data at hyperscale has not been publicly demonstrated. Additionally, the integration of 3D photonic interconnects may limit flexibility for customers who prefer modular, hot-swappable optical modules. Pricing is not disclosed, but the technology targets the high-end AI cluster and data center switch market, likely commanding a premium over conventional optics.
How it works
-
Edgeless I/O architecture
Places optical I/O across the entire die area, enabling bandwidth density to scale with area (mm²) rather than perimeter.
-
3D photonic interconnect
Embeds fiber-optic connections directly beneath computer processors, eliminating shoreline bandwidth bottlenecks.
-
L200 co-packaged optics
Supports 32 to 64 Tbps aggregate bandwidth using 112G PAM4 signaling for dense, fast interconnects in AI training clusters.
-
L20 high-density module
Delivers 12.8 Tbps aggregate bandwidth at 4x pluggable density, enabling high-capacity deployments in compact systems.
-
M1000 3D interposer
A rack-level reference platform delivering 114 Tbps across a 4,000 mm² footprint for frontier-scale training.
-
EVK100 bidirectional link
Provides up to 3.2 Tbps per fiber at 1.9 pJ/bit, demonstrating high-bandwidth, energy-efficient bidirectional photonic links.
-
Guide VLSP light engine
Generates up to 51.2 Tbps per laser module with 100 mW per fiber across 16 wavelengths, enabling hyperscale CPO deployments.
Strengths and trade-offs
Strengths
- Edgeless I/O architecture allows bandwidth density to scale with die area, not perimeter, providing unlimited scaling headroom for future chips.
- The Passage M1000 EVK delivers 114 Tbps across a 4,000 mm² footprint, demonstrating rack-level 3D photonic interposer capability.
- The Guide light engine enables 100 Tbps of switch bandwidth in a compact 1RU chassis, replacing about 18 conventional ELSFP modules that would occupy 4RU.
- The EVK100 achieves 3.2 Tbps per fiber at 1.9 pJ/bit, setting a benchmark for energy-efficient bidirectional photonic links.
Trade-offs
- Passage is only available for early access partners, not yet in mass production, limiting immediate deployment for hyperscalers.
- The platform requires significant infrastructure changes (new packaging, cooling, fiber management) compared to pluggable optics, increasing initial deployment complexity.
- Field reliability data at hyperscale has not been publicly demonstrated, despite claims of superior yields and reliability from foundry-style production.
- 3D photonic integration may reduce flexibility for customers who prefer modular, hot-swappable optical modules in their existing architectures.
Pricing context
Not publicly disclosed; available for early access partners with custom pricing expected for hyperscale AI clusters and data center switches.
Getting started with Passage
-
Request early access
Visit the Lightmatter website and submit an early access request for the Passage platform. Provide your organization details, use case, and expected deployment scale to initiate the evaluation process.
-
Integrate Guide light engine
Connect the Guide VLSP light engine to your system using the provided fiber-optic cabling. Ensure the laser module is properly aligned and powered to deliver up to 51.2 Tbps per module for your photonic interconnects.
-
Configure Passage module
Select the appropriate Passage form factor (e.g., L200, L20, or M1000) based on your bandwidth needs. Mount the module onto the processor or interposer, and connect the optical fibers to the designated ports.
-
Run bidirectional link test
Use the EVK100 or EVK50 evaluation kit to establish a bidirectional photonic link. Verify the link achieves the specified bandwidth (up to 3.2 Tbps per fiber) and energy efficiency (1.9 pJ/bit) using the provided testing software.
-
Deploy in AI cluster
Integrate the Passage-enabled switches and processors into your AI training cluster. Monitor the interconnect performance using Lightmatter's management tools, and scale the deployment by adding more modules as needed.
Frequently Asked Questions
What is Passage and how does it work?
Passage is a 3D photonic interconnect platform from Lightmatter that embeds fiber-optic connections directly beneath processors. Its Edgeless I/O architecture places optical I/O across the entire die area, allowing bandwidth density to scale with area rather than chip perimeter.
What are the key products in the Passage lineup?
Passage includes the L200 with 32 to 64 Tbps aggregate bandwidth, the L20 delivering 12.8 Tbps at 4x pluggable density, the M1000 EVK providing 114 Tbps across a 4,000 mm² footprint, and the EVK100 and EVK50 bidirectional links with up to 3.2 Tbps per fiber.
How does the Guide light engine support Passage interconnects?
The Guide light engine uses Very Large Scale Photonics technology to generate up to 51.2 Tbps per laser module. It provides a minimum of 100 mW per fiber across 16 multiplexed wavelengths with precise 200 GHz spacing, enabling bidirectional photonic links for hyperscale co-packaged optics deployments.
What are the main benefits of Passage's Edgeless I/O architecture?
Edgeless I/O places optical I/O across the full die area, so bandwidth density scales with area (mm²) rather than perimeter. This eliminates the shoreline bottleneck of traditional electrical I/O, providing unlimited scaling headroom for future chips and massive bandwidth density for AI training clusters.
How does Passage compare to other photonic interconnect solutions?
Passage competes with Ayar Labs, Intel silicon photonics, and Marvell PAM4 DSPs. Its key differentiator is the 3D Edgeless I/O architecture and Guide light engine, which reduce component counts and assembly complexity compared to discrete ELSFP modules, enabling higher density in smaller form factors.
Is Passage available for purchase and what are the deployment considerations?
Passage is currently available only for early access partners, not in mass production. It requires significant infrastructure changes like new packaging, cooling, and fiber management compared to pluggable optics. Pricing is not publicly disclosed, and field reliability data at hyperscale has not been demonstrated.
Alternatives
How Passage compares
Direct head-to-head against 2 competitors. Picked by 7wData.
Passage
- Pricing
- Not publicly disclosed; available for early access partners with custom pricing expected for hyperscale AI clusters and data center switches.
- Target
- Passage is a 3D photonic interconnect platform from Lightmatter, designed to eliminate the bandwidth bottleneck in AI training and data center switching.
- Strength
- Edgeless I/O architecture allows bandwidth density to scale with die area, not perimeter, providing unlimited scaling headroom for future chips.
- Watch for
- Passage is only available for early access partners, not yet in mass production, limiting immediate deployment for hyperscalers.
Eventbrite
- Pricing
- No pricing found
- Target
- Event management and ticketing
- Deployment
- Cloud-based
- Strength
- Branded checkout and analytics
- Watch for
- Limited customization for niche markets
Purplepass
- Pricing
- No pricing found
- Target
- Enterprise ticketing
- Deployment
- Cloud-based
- Strength
- Free seat maps and 24/7 US support
- Watch for
- Higher cost for advanced features
User reviews
No user reviews yet. Be the first to write one.
Sources
Reporting on this tool draws on these publicly available sources.