A New Frontier for Edge AI Computing

Singapore-based startup Acrab has introduced the Agent Box, a compact desktop system engineered to handle 100-billion-parameter language models directly on local hardware. By moving AI workloads away from the cloud, the device aims to enhance privacy, reduce operational costs, and ensure consistent performance even without a stable internet connection.


The Power of the G≡LIX 1 Processor

At the heart of the Agent Box lies the custom-designed G≡LIX 1 chip, which utilizes a 5nm manufacturing process to optimize edge computing. The processor features a sophisticated architecture, including:

  • A 20-core Arm CPU.
  • A 3 TFLOPS-rated GPU.
  • A specialized multicore neural processing unit (NPU) tailored for large language model (LLM) inference.
  • High-bandwidth unified memory reaching 273GB/s, complemented by an 8MB L1 cache and a 768GB/s L2 cache.

Acrab reports that this chip achieves 700 TOPS of processing power. Furthermore, the integration of VAE vector acceleration reportedly improves attention processing speeds by up to 300%.


Performance Benchmarks and Real-World Potential

According to internal company data, the Agent Box demonstrates remarkable efficiency. During testing with a Gemma 26B configuration, the system achieved a prefill rate of 1,416.8 tokens per second. In comparison, standard consumer hardware like the Apple Mac Mini M4 Pro recorded 188.9 tokens per second, indicating that Acrab's solution offers a substantial speed advantage for specific AI tasks.


«Generative AI helped people find answers. Agentic AI will help them get things done,» stated Dr. Ken Phua, CEO of Acrab.

Expanding the Agentic AI Ecosystem

The Agent Box is more than just hardware; it includes a comprehensive software suite comprising developer toolchains, AI runtimes, and orchestration software for autonomous agents. Early demonstrations have showcased the device's ability to execute voice commands that control smart home appliances, operate robotics, and generate 3D printing files.


Looking ahead, Acrab plans to integrate its computing platform into various sectors, including industrial robotics, smart vehicles, and AI-enabled Network Attached Storage (NAS) systems. While details regarding the company’s internal operations remain limited, the project has drawn interest due to potential ties to the major battery manufacturer CATL, which appears in the project's copyright documentation.


«Our goal is to give device makers and developers the foundation to bring agentic intelligence into many different products and environments,» added Dr. Phua regarding the company's broader vision for the GΞLIX 1 ecosystem.