Rebellions REBEL-Quad Champions a Chiplet-First Strategy for Large-Scale AI Inference

Rebellions REBEL‑Quad targets efficient large‑scale inference with a chiplet design, high‑bandwidth links, and FP8/FP16 performance under 600 W.

Inference at scale relies on power efficiency as much as raw performance. AI datacenters and hyperscalers want predictable, low-power performance while executing long-running, memory-dominated workloads. Rebellions REBEL-Quad aims to address these challenges through a device-local execution model that embeds task scheduling, synchronization, and data movement within the accelerator fabric rather than relying on host-driven orchestration. REBEL-Quad is a chiplet-based inference accelerator encompassing four homogeneous compute dies coupled via a high-bandwidth UCIe-Advanced die-to-die mesh. The interconnect supports up to 1 TB/s per channel and delivers mixed-precision inference using FP8 and FP16. As part of this, the chip delivers up to 2 and 1 PFLOPS of FP8 and FP16 throughput, respectively, while keeping power consumption under 600 W. Rebellions pairs this hardware structure with a PyTorch-native software framework that integrates graph-mode optimization, vLLM-based serving, and precision-aware execution.