Fujitsu AI Computing Broker (ACB)

Runtime-aware GPU resource management middleware for AI workloads

What is ACB?

Fujitsu AI Computing Broker (ACB) is an intelligent middleware that dynamically optimizes GPU allocation and manages memory oversubscription for AI/ML workloads. It monitors your AI framework in real-time to allocate GPUs only when needed, maximizing efficiency and reducing computing costs.

Key Benefits

• Improved GPU Utilization: Automatically reclaims idle GPU time during CPU-bound phases (preprocessing, data loading)
• Higher Throughput: Advanced scheduling with backfill optimization enables running more jobs concurrently
• Zero Code Changes: Deploy with automatic mode - no modification to existing PyTorch programs required
• Memory Oversubscription: Run multiple models on a single GPU when memory allows
• Cost Reduction: Get more value from your GPU infrastructure through better resource utilization

Core Features

• Runtime-Aware Allocation: Monitors PyTorch framework to assign GPUs dynamically during compute-intensive phases
• Full Memory Access: Active programs get complete GPU memory without virtualization constraints (unlike MIG/vGPU)
• Intelligent Scheduling: FIFO, GPU-sharing, and GPU-affinity schedulers with backfill support
• Multi-GPU & Multi-Node: Supports distributed training with PyTorch DDP across multiple nodes
• Fast Deployment: Works with unmodified programs via automatic interception or manual API integration

Quick Start

# Install
pip install ai-computing-broker

# Start GPU Assigner
gpu_assigner start

# Run your existing PyTorch program (no code changes!)
AGA_ENABLE_AUTO=1 agarun python your_training_script.py