English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

Pre-Compiled Pipeline Shards for Distributed LLM Inference on Intel AI PCs

Forum topic · 小凯 · 2026-08-21

Summary

A new arXiv paper (2608.19147) by Tate Berenbaum and Muthaiah Venkatachalam shows that a handful of Intel AI PCs, connected over an ordinary network, can serve LLMs too large for any single device. The system uses pipeline parallelism: the model is split by layer into per-stage shards, each pre-compiled as an OpenVINO graph, with each machine running one shard and passing activations to the next. Three techniques make the approach practical: injecting a beam_idx Gather into each shard to trigger OpenVINO's IndirectKVCache fusion and recover monolithic-level speed; speculative decoding on stateful OpenVINO models; and micro-batching that interleaves concurrent requests, each carrying its own KV cache. A two-node Llama 3.1 8B INT4 pipeline served two concurrent users at 1.79x the throughput of a single-user unsplit model on identical hardware, with the gap widening under simulated WAN latency. A four-node Lunar Lake AI PC deployment on Intel Tiber Cloud served a 70B model at interactive speed, with outputs token-for-token identical to non-speculative decoding. Code, benchmark logs, and reproduction scripts are available on GitHub.

Paper Overview

Field: Machine Learning Authors: Tate Berenbaum, Muthaiah Venkatachalam Published: 2026-08-19 arXiv: 2608.19147 Code: github.com/labscommunity/pipeline-sharded-inference-paper (reproduction scripts in the top-level reproduction/ directory)

Summary

Modern Intel AI PCs ship capable integrated GPUs and NPUs with 16+ GB of unified memory, yet they spend considerable time idle. That memory is not enough to fit large models such as a 70B-parameter LLM. This paper demonstrates that a handful of AI PCs working together over an ordinary network can serve models beyond the capability of any single device.

The system uses pipeline parallelism: the model is split by layer into per-stage shards, each pre-compiled into an OpenVINO graph, so every machine runs one shard and passes activations to the next.

Three Key Techniques

1. Recovering unsplit-model speed: A naive per-stage export runs well below monolithic inference because it misses an OpenVINO GPU optimization. Injecting a beam_idx Gather into each shard triggers the IndirectKVCache fusion and brings shards to parity with the unsplit model.

2. Speculative decoding on stateful OpenVINO models: The authors adapt speculative decoding to work with OpenVINO's stateful inference.

3. Micro-batching: The pipeline serves multiple concurrent users by interleaving requests across stages, with each request carrying its own KV cache.

Results

  • A two-node Llama 3.1 8B INT4 pipeline served two concurrent users at 1.79x the throughput of a single-user unsplit model on identical hardware; the gap widened under simulated WAN latency.
  • The same design scaled to a 70B model that no single cluster member could hold: a four-node Lunar Lake AI PC deployment on Intel Tiber Cloud served a single user at interactive speed, with outputs token-for-token identical to non-speculative decoding of the same four-node pipeline.
Code, raw benchmark logs, and reproduction scripts are provided as a self-contained package at https://github.com/labscommunity/pipeline-sharded-inference-paper.

*Collected automatically on 2026-08-21.*

Tags

#llm-inference#intel-ai-pc#pipeline-parallelism#openvino#speculative-decoding#distributed-computing#arxiv

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178633743