Lightmatter expanded last year's analysis of training and shows meaningful benefits of CPO for inference systems.
Public source
Publisher name
Public post
Congratulations to Arul for leading this paper effort! This expands last years analysis of training and shows meaningful benefits of CPO for inference systems.
Company
Lightmatter
A Giant Leap
- Industry
- Computer Hardware Manufacturing
- Location
- Mountain View, US
- Company size
- 201–500 employees
About Lightmatter
We are building the engine that will power discoveries, drive progress, and redefine what computers—and human beings—are capable of.
See moreLatest activity
Latest activity from Lightmatter
7 signals
Products & Services
Lightmatter released the Passage L-Series CPO chip delivering up to 64 Tbps bandwidth to enable 3x faster training and lower latency inference.
Research & Knowledge
Lightmatter highlights that photonic interconnects will allow data centers to join GPUs in a rack or entire ROW of racks to form a supercomputer with 1,024 or more GPUs within a matter of years.
Presence & Recognition
Lightmatter Founder and CEO Nicholas Harris is featured in the AI Infra Summit first panel titled Feeding the AI Factory: Solving Data Movement, Memory, and Storage Bottlenecks at Scale.
Discover more
Similar signals
Similar public activity from other companies.
Research & Knowledge
Efficient Computer
Efficient Computer published research on the Electron E1 general-purpose processor, where the int8 matrix multiply runs at 18.3x less energy than an ARM Cortex-M4F based SoC.
Research & Knowledge
Tensordyne
Tensordyne published a whitepaper on its approach to solving AI inference challenges.
Research & Knowledge
Lumai
Lumai built Lumai Iris to address the constraint that matrix multiplication accounts for well over 90% of the arithmetic in a transformer inference call, with compute and energy being coupled.
Research & Knowledge
SambaNova
SambaNova published a blog post detailing Raghu Prabhakar's insights from the Hot Chips Symposium 2026, including how SN50 is designed to keep data moving efficiently as models scale.
Research & Knowledge
Extropic