OpenVector
  • OpenVector
  • Introduction
    • Our Thesis
    • Our Solution
    • Other Approaches
    • Benchmarks
  • Network Architecture
    • Overview
    • Networking
    • Deploy Compute Node
    • Deploy CoFHE Node
    • Deploy Client Node
  • Quick start
    • Overview
    • Setting up Client Node
    • Encrypting Data
    • Tensor Multiplication
    • Decrypting the Ouput
    • Verifying the Output
    • Running the Program
  • Tutorials
    • Building MLP Block
  • Use Cases
    • Confidential LLM Inference
    • Training on Encrypted Data
    • Vector Search on Encrypted Data
    • Encrypted Data Access Control
  • API references
    • CryptoSystem Interface
    • ClientNode Class
    • ComputeRequest Class
    • ComputeResponse Class
  • PYTHON API REFERENCES
    • Overview
    • Tensor
    • nn.Module
    • nn.Functional
  • Contribution
    • Call for Research
    • CoFHE Library
Powered by GitBook
On this page
  • Subnet A: High-Speed Connectivity with InfiniBand
  • Subnet B: Standard Internet Connectivity for Scalability
  1. Network Architecture

Networking

The OpenVector network employs distinct networking technologies tailored to the requirements of its two subnet types: Subnet A for low-latency applications and Subnet B for non-latency-critical use cases.

Subnet A: High-Speed Connectivity with InfiniBand

Subnet A leverages InfiniBand, a leading implementation of Remote Direct Memory Access (RDMA), to enable high-speed, low-latency networking. InfiniBand is specifically designed for high-performance computing and data center environments, making it an ideal choice for latency-critical operations such as inference.

Subnet B: Standard Internet Connectivity for Scalability

Subnet B employs standard internet connectivity, prioritizing accessibility and scalability over low-latency performance. This configuration is well-suited for non-latency-critical use cases, such as decentralized training, where real-time responsiveness is not a primary concern. While Subnet B does not achieve the ultra-low latency of InfiniBand, its design ensures that OpenVector remains accessible to a broader range of participants, balancing cost, scalability, and functionality for non-latency-sensitive computations

PreviousOverviewNextDeploy Compute Node

Last updated 3 months ago