Specifications

Overview

The Rebellions REBEL (RBLN-CR13) is a chiplet-based AI accelerator for data center inference and training, fabricated by Samsung on a 4 nm process. Each die pairs 16 neural cores with 128 MB of software-managed on-die memory, and the card carries 144 GB of HBM3E running at 4,912 GB/s — the memory capacity and bandwidth tier that frontier LLM serving actually depends on.

What distinguishes REBEL architecturally is its interconnect. Alongside a PCIe Gen5 x16 host link (128 GB/s), the part exposes three UCIe-A chip-to-chip ports at 1 TB/s each, so accelerators can be composed into REBEL-Quad nodes with a much higher-bandwidth, lower-latency inter-device fabric than PCIe-based scale-out provides. Compute is rated at 1,024 TFLOPS FP16 and 2,048 TFLOPS in FP8/FP4 formats at a 350 W TDP.

Rebellions targets Korean and APAC hyperscale and sovereign-AI deployments with REBEL, and the RBLN software stack provides the compiler and runtime for model deployment. QS Compute can source RBLN-CR13 accelerators and REBEL-Quad systems for enterprise buyers — contact us for configuration, pricing and lead time.