REBEL-Quad is an advanced AI SoC built on a UCIe-Advanced chiplet architecture, designed to meet the extreme compute and memory demands of serving frontier LLMs at massive scale for hyperscalers, AI data centers, and enterprises. It features a unified hardware-software stack that maximizes utilization and delivers exceptional performance-per-watt across both compute-bound prefill and memory-bound decoding phases. The chiplet-based design enables hyper scalability without compromising latency or coherence.
Key enablers include
- Unified Mixed Precision Compute Engine
- Predictive DMA & On-Chip Mesh
- Holistic Synchronization mechanism
- UCIe-Advanced and Amba Chi C2C
REBEL-Quad’s modular design is enabled by the AMBA CHI C2C protocol running on UCIe-Advanced physical link. This technology is used to connect the Neoverse Compute Subsystem based chiplet with the REBEL-Quad and provides a path to next generation AI accelerators. Through this chiplet expansion, we can build NVIDIA’s Grace Hopper-like AI accelerator with a powerful Arm CPU chiplet & AI Accel all in one package. Rebellions participates in Arm-led Chiplet System Architecture (CSA) collaboration to ensure that the Rebel-Quad is compliant with Arm CSS designs.
Rebellions is also a proud member of the Arm Total Design (ATD) Program, working closely with industry experts to promote standards and progress custom AI design.
