Skip to content

Home

Fujitsu AI computing broker
Fujitsu AI computing broker (ACB) is a run-time aware middleware that optimizes GPU allocation and manages memory oversubscription. The result: improved efficiency, higher throughput, and reduced computing costs.

Your GPUs are not lazy —Your Scheduler is:
ACB makes GPUs work harder — not cost more

figure1
Key Features

● Runtime-aware GPU allocation: Monitors AI framework to allocate GPUs needed
● Full Memory Access: Active program has access to the full GPU memory
● Advanced Scheduling: Employs techniques like backfill to optimize job placement and maximize aggregate utilization
● Fast Deployment: No code changes required in user programs

Unlock all the power your GPUs can offer, free, for 30 days

figure2
Your Early Access Starts Here Experience the performance, flexibility, and control of our solution-free, with no commitment.
Want to Dive Deeper? Check out the Walking Deck, Technical White Paper, and additional resources at the bottom of this page.
Questions? Reach us by clicking the "Early Access: Learn More Here – Free Beta" button on this page.

Do any of the following apply to you?

  1. Your workloads involve alternating phases of CPU and GPU tasks
    		2.
  2. You have multiple GPU-capable jobs that can be queued or batched
    		3.
  3. Your workloads can utilize full GPU memory intermittently
    		4.
  4. You need to host multiple domain-specific LLMs with uneven demand on shared GPU infrastructure
    		5.

Technology Overview

Target Industry/Users

AI Infrastructure teams, Enterprise ML and Data teams, Cloud/SaaS AI Service providers, AI Research labs and Startups (Biotech, Finance, Analytics), GPU Hosting Providers

Challenges in Target Industry and Operations

Soaring prices and shortages of GPUs, underutilization of current GPU infrastructure

Technical Challenges

Runtime inefficiencies, fragmented toolchains, hardware scaling constraints

Fujitsu's Technological Advantage

  • Runtime aware GPU allocation
  • GPU memory oversubscription management
  • Integration without code changes in the user programs
  • See technical white paper for a detailed introduction (linked at the bottom)

Use Cases

  • LLM inference
    • Installing multiple AI applications on a single inference server to provide diverse services
  • AI Training and Inference
    • Improving development throughput by enabling simultaneous execution of multiple learning processes during model parameter exploration

Case Studies

  • AlphaFold2 Throughput Optimization
    • ACB dynamically shares GPU resources between multiple AlphaFold2 jobs, maintaining throughput increasing GPU utilization by 45%
  • Multi-LLM Hosting
    • ACB manages multiple LLMs on a single GPU server, optimizing memory usage and enabling concurrent model serving without delays
  • FX Risk Models
    • Improved training throughput for financial risk models by multiplexing GPU usage, enhancing efficiency by over 2-fold
  • Cloud Services
    • Host several AI models on a single GPU node and manage five times the physical GPU memory capacity
  • AI models for Retail
    • Achieved 2× throughput in object recognition model training for IaaS applications

(see press release and technical white paper linked at the bottom for more details)

Technical Trial

  • Join Private Beta: Get 30 days of full access. Test, explore, and shape what is next
  • A Proof of Concept is possible

Documents (coming soon)

English vesion Japanese version
ACB Walking Deck ACB Walking Deck
Technical White Paper Technical White Paper
User Manual User Manual