VS4x400 is a high-performance, dual-core hardware accelerator for vector similarity search, designed for next-generation on-device AI.

---

🔹 Overview

VS4x400 enables ultra-fast similarity search for embeddings on edge devices. Its architecture allows low-latency vector computations while maintaining high energy efficiency. Ideal for embedding-based search, recommendation systems, and real-time AI applications.

---

⚙️ Features

• Dual-core accelerator optimized for vector operations.
• Zero-overhead hardware pipeline for N×D matrix-vector computations.
• Support for large embeddings (up to 16k×1024 tested).
• Energy-efficient design with extreme speedups compared to CPU implementations.
• Easy integration into existing Python or C++ pipelines.

---

🏎️ Benchmark Specs

We evaluated the VS4x400 dual-core vector accelerator using different database sizes and embedding dimensions. Results show significant speedups and energy efficiency compared to software execution on a standard CPU.

Config (N×D)	Hardware Cycles	Hardware Time (ms)	Software Time (ms)	Speedup	Energy Gain
1024 × 128	16,384	0.0410	0.1415	3.45×	345.4×
4096 × 256	131,072	0.3277	1.4109	4.31×	430.6×
8192 × 512	524,288	1.3107	6.3500	4.84×	484.5×
16,384 × 1024	2,097,152	5.2429	24.0144	4.58×	458.0×

Note: Hardware time is estimated for the VS4x400 running at 400 MHz with 8 MAC cores. Energy gain is calculated relative to a CPU consuming 10 W versus the VS4x400 consuming 0.1 W.

---

🛠️ Installation

# Clone the repository
git clone https://github.com/johanlabs/vs4x400.git
cd vs4x400

# Install dependencies
yarn install
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

---

⚡ Quick Start

Run the benchmark script to verify performance:

yarn benchmark

Example output:

=== VS4x400 Zero-Overhead Benchmark ===
Config: N=1024, D=128
  Hardware Time:   0.0410 ms
  Software Time:   0.1415 ms
  SPEEDUP:         3.45x
  ENERGY GAIN:     345.4x
...

---

📚 Usage

Python API

from vs4x400 import VectorR

engine = VectorR()
embeddings = engine.load_embeddings("data/embeddings.npy")
results = engine.query(embeddings, query_vector)

C++ API

#include "vs4x400.hpp"

VectorR engine;
engine.load_embeddings("data/embeddings.bin");
auto results = engine.query(query_vector);

---

📄 License

This project is licensed under the ISC License – see the LICENSE file for details.