# OBM Distributes AI Weights Over Optics, With Up to 82% HBM Savings Estimated Across 80,000 Accelerators

DevFeed: [OBM Distributes AI Weights Over Optics, With Up to 82% HBM Savings Estimated Across 80,000 Accelerators](<https://devfeed.tech/articles/obm-distributes-ai-weights-over-optics-with-up-to-82-hbm-savings-estimated-across-80-000-accelerators-134017.md>)

Original publisher: [Read original article](<https://joho-todai.com/en/obm-optical-bandwidth-memory-hbm-reduction/>)

Author: Toru Misaki

Published: 2026-11-14T11:12:40Z

Content type: article

Language: en

Sources: [Joho no Todai](<https://devfeed.tech/sources/joho-no-todai.md>)

Topics: [Neural Network](<https://devfeed.tech/topics/neural-network.md>), [quantization](<https://devfeed.tech/topics/quantization.md>), [communications](<https://devfeed.tech/topics/communications.md>)

Tags: [2026](<https://devfeed.tech/tags/2026.md>), [accelerators](<https://devfeed.tech/tags/accelerators.md>), [ai](<https://devfeed.tech/tags/ai.md>), [broadcast](<https://devfeed.tech/tags/broadcast.md>), [central](<https://devfeed.tech/tags/central.md>), [configuration](<https://devfeed.tech/tags/configuration.md>), [fiber](<https://devfeed.tech/tags/fiber.md>), [hbm](<https://devfeed.tech/tags/hbm.md>), [memory](<https://devfeed.tech/tags/memory.md>), [model](<https://devfeed.tech/tags/model.md>), [model-weights](<https://devfeed.tech/tags/model-weights.md>), [optical](<https://devfeed.tech/tags/optical.md>), [optics](<https://devfeed.tech/tags/optics.md>), [over](<https://devfeed.tech/tags/over.md>), [simulations](<https://devfeed.tech/tags/simulations.md>), [source](<https://devfeed.tech/tags/source.md>), [systems](<https://devfeed.tech/tags/systems.md>), [university](<https://devfeed.tech/tags/university.md>), [up](<https://devfeed.tech/tags/up.md>), [with](<https://devfeed.tech/tags/with.md>)

## AI overview

Optical Broadcast Memory (OBM) sends shared, read-only model weights from a central server over fiber to accelerators, leaving variable data such as the KV cache in local HBM. Simulations across 80,000 accelerators estimate HBM reductions of 39-40% for Llama 3.1 405B and Llama 4 Maverick, with throughput near baseline; DeepSeek V4 Pro reaches 82% savings. Its reported 1.57x throughput depends strongly on baseline memory capacity and placement: at 192 GB HBM per accelerator, the ratio falls to 0.96x. The study covers decode only, and its multicore optical amplifier remains conceptual. It also identifies a workload skew in which one popular MoE expert can no longer keep processing aligned with weight arrivals; dynamic expert replication is left for future work.

## Source excerpt

Uppsala University has proposed Optical Broadcast Memory (OBM), which distributes model weights copied across tens of thousands of systems from a central source over optical fiber. Simulations of an 80,000-accelerator configuration estimate HBM savings of up to 82%.