inference/research
verifiedhigh · 0.90updated 2026-10-11

Lightmatter Passage L20 Photonic Interconnect

800 GB/s

Fields

id
interconnect/lightmatter-passage-l20
Kind
other
Version
Passage L20 (212.5 Gbps PAM4, 32 data fibers)
Bandwidth (GB/s, per link, unidirectional)
800
Bandwidth basis
PER L20 MODULE, PER DIRECTION. Lightmatter's technical specifications table prints verbatim: 'Bandwidth: 6.4 Tbps Each direction', alongside 'SerDes line rate: 212.5 Gbps PAM4' and 'Data fibers: 32 Corning CPO Flexconnect'. 6.4 Tbps per direction converts to 6,400 / 8 = 800 GB/s in each direction per L20 module. Per data fiber, this corresponds to 6.4 Tbps / 32 fibers = 200 Gbps (= 25 GB/s per fiber per direction). Bidirectional aggregate bandwidth per L20 module is 12.8 Tbps (= 1,600 GB/s bidirectional aggregate).
Link count
32
Topology
Near-Packaged Optics (NPO) on host PCB/mezzanine near ASIC, or On-Board Optics (OBO) with retimer. Interfaces to optical fabric via 32 detachable Corning CPO Flexconnect fibers.
Scale-up
yes
Scale-out
yes
Switching
Passive optical engine with built-in optical circuit switching capability (Passage platform). Driven directly by 212.5 Gbps PAM4 SerDes; packaged in an 1827-ball BGA reflow-compatible footprint with cold-plate liquid cooling.
notes
Lightmatter Passage L20 is a high-density BiDi (bidirectional) photonic interconnect module targeting Near-Packaged Optics (NPO) and On-Board Optics (OBO) deployments. Unlike conventional pluggable optical transceivers (QSFP-DD/OSFP) that crowd the chassis faceplate and require long PCB copper traces with power-hungry retimers, the L20 sits on the PCB immediately adjacent to host XPUs or switches.\n\nINFERENCE RELEVANCE: Inter-node communication is the primary throughput and cost bottleneck in distributed LLM inference across multi-node clusters. During autoregressive decode of large models (such as 70B+ or MoE architectures partitioned across nodes via tensor or expert parallelism), latency is dominated by collective all-reduce and all-to-all communications. Pluggable transceivers introduce substantial electrical SerDes loss and thermal limits. Passage L20 delivers 6.4 Tbps per direction per module at an ultra-low energy profile with zero faceplate transceiver bottleneck, allowing tight rack-scale optical interconnect topologies that drastically reduce tail latency and power in distributed inference fabrics.
way
each

Notes

Lightmatter Passage L20 is a high-density BiDi (bidirectional) photonic interconnect module targeting Near-Packaged Optics (NPO) and On-Board Optics (OBO) deployments. Unlike conventional pluggable optical transceivers (QSFP-DD/OSFP) that crowd the chassis faceplate and require long PCB copper traces with power-hungry retimers, the L20 sits on the PCB immediately adjacent to host XPUs or switches.\n\nINFERENCE RELEVANCE: Inter-node communication is the primary throughput and cost bottleneck in distributed LLM inference across multi-node clusters. During autoregressive decode of large models (such as 70B+ or MoE architectures partitioned across nodes via tensor or expert parallelism), latency is dominated by collective all-reduce and all-to-all communications. Pluggable transceivers introduce substantial electrical SerDes loss and thermal limits. Passage L20 delivers 6.4 Tbps per direction per module at an ultra-low energy profile with zero faceplate transceiver bottleneck, allowing tight rack-scale optical interconnect topologies that drastically reduce tail latency and power in distributed inference fabrics.

Cross-references

This record references no other records — it is a leaf. It is cited by 0.