Gespeichert in:
Bibliographische Detailangaben
1. Verfasser: Momen Ghazouani
Format: Recurso digital
Sprache:Englisch
Veröffentlicht: Zenodo 2026
Online-Zugang:https://doi.org/10.5281/zenodo.18130207
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866901537630978048
author Momen Ghazouani
author_facet Momen Ghazouani
contents <p>The training of large-scale neural networks increasingly suffers from communication bottlenecks across distributed accelerators, with gradient synchronization dominating both bandwidth consumption and energy costs. Conventional parameter exchange protocols operate under a strictly explicit transmission paradigm: every update, regardless of magnitude or novelty, is broadcast in full across the cluster. Empirical profiling of modern large language model (LLM) training reveals that 60–75% of gradient shards exhibit high temporal redundancy changes that fall below optimizer-effective thresholds or maintain near-identical values across consecutive iterations. Yet current all-reduce pipelines expend identical bandwidth on both critical and negligible updates, leading to disproportionate energy drain and interconnect congestion We introduce the Silent Gradient Protocol (SGP), a silence-aware training communication framework that reframes absence of signal as a valid synchronization state. Unlike traditional collectives that equate non-transmission with failure, SGP establishes a formal contract where intentional silence denotes that the accumulated gradient update is statistically negligible. Consequently, the receiver interprets silence as a zero-vector update, while the sender locally preserves the residual information for future iterations By elevating silence to a first-class protocol primitive, SGP eliminates redundant gradient traffic while maintaining convergence guarantees. Our deterministic distributed Simulation demonstrates that SGP achieves 90–95% communication reduction (Global Compression Ratio) while preserving model convergence within ±0.15% of dense baseline accuracy, effectively decoupling communication cost from model size in distributed neural training </p>
format Recurso digital
id zenodo_https___doi_org_10_5281_zenodo_18130207
institution Zenodo
language eng
publishDate 2026
publisher Zenodo
record_format zenodo
spellingShingle Silence Is All We Need
Momen Ghazouani
<p>The training of large-scale neural networks increasingly suffers from communication bottlenecks across distributed accelerators, with gradient synchronization dominating both bandwidth consumption and energy costs. Conventional parameter exchange protocols operate under a strictly explicit transmission paradigm: every update, regardless of magnitude or novelty, is broadcast in full across the cluster. Empirical profiling of modern large language model (LLM) training reveals that 60–75% of gradient shards exhibit high temporal redundancy changes that fall below optimizer-effective thresholds or maintain near-identical values across consecutive iterations. Yet current all-reduce pipelines expend identical bandwidth on both critical and negligible updates, leading to disproportionate energy drain and interconnect congestion We introduce the Silent Gradient Protocol (SGP), a silence-aware training communication framework that reframes absence of signal as a valid synchronization state. Unlike traditional collectives that equate non-transmission with failure, SGP establishes a formal contract where intentional silence denotes that the accumulated gradient update is statistically negligible. Consequently, the receiver interprets silence as a zero-vector update, while the sender locally preserves the residual information for future iterations By elevating silence to a first-class protocol primitive, SGP eliminates redundant gradient traffic while maintaining convergence guarantees. Our deterministic distributed Simulation demonstrates that SGP achieves 90–95% communication reduction (Global Compression Ratio) while preserving model convergence within ±0.15% of dense baseline accuracy, effectively decoupling communication cost from model size in distributed neural training </p>
title Silence Is All We Need
url https://doi.org/10.5281/zenodo.18130207