/
Work stream 9: Deterministic Latency Computing Technology

Work stream 9: Deterministic Latency Computing Technology

  • Leader: @Jun Ogawa

  • Objective: Establish a deterministic latency computing platform that enables real-time AI processing of periodic data streams across edge and data center, while bounding processing jitter and improving energy efficiency.

  • Approach: Develop and validate a deterministic latency computing platform based on the OSS (release planned by Mar. 2027), incorporating time-slot-based pipeline planning and runtime execution.

1. Introduction

  • Basic Concept of the Deterministic Latency Computing Platform

    • Expands deterministic latency from optical networking to computing systems.

    • Enables real-time AI inference and analytics of periodic data streams such as video.

    • Balances low power consumption with real-time performance.

  • Why InfiniEdge AI?

    • As our concept aligns with the goals of InfiniEdge AI, we hope to contribute through a new Work stream.

    • A valuable community for gathering feedback on our ideas and software and for seeking potential collaborations.

fig1.jpg

2. Edge Computing vs. Data-Center Computing

  • Edge computing bounds processing jitter but limits aggregation efficiency.

  • Centralized data centers provide better aggregation efficiency, but computing tasks there introduce jitter in both data transmission and processing.

fig2.jpg

3. Challenges

  • Existing technologies improve aggregation efficiency while reducing jitter in transmission.

    • Optical network provides a deterministic transmission latency between edge sites and a data center.

  • Our challenge: Bounding processing jitter within a defined time window in a data center for expanding deployment flexibility of tasks across edge and data center.

fig3.jpg

4. Concept Overview

fig4.jpg

5. Use Case: Collision Risk Assessment

  • Assessing and predicting vehicle collision risks at intersections to provide real-time feedback to drivers.

    • Integrating and analyzing data from multi-view cameras.

  • Enabling data aggregation across multiple intersections with both real-time performance and low power consumption.

a4aab1d5-2945-4fb7-84e6-665a5c82a71d.jpg

6. Target Model

  • Periodic data arrival.

  • Latency targets: tens to hundreds of milliseconds.

fig6.jpg

 

 

 

 

 

 

7. Architecture for Deterministic Latency Computing

  • Logical data pipeline planning (Controller) by NTT

    • Time-slot-based data pipeline to ensure deterministic processing for multiplexed data streams.

    • Logical task planner to generate task-allocation plans and Pipeline optimizer for mapping the plans onto data-center resources.

  • Pipeline execution (Runtime) by 1FINITY

    • Time-slot-aligned task execution and termination mechanism helps ensure that processing completes within the required time window.

    • Lock-free memory-access mechanism synchronized with time slots enables low-jitter memory access.

fig7.jpg

 

Slides used for the proposal on July 29, 2026