Rebellions REBEL-Quad
Vendor: Rebellions (South Korea)
Category: ASIC Dedicated Accelerator
Architecture: UCIe-Advanced Chiplet
Introduction
Rebellions REBEL-Quad is a data center-grade AI accelerator card from South Korean AI chip company Rebellions, based on a UCIe-Advanced 4-chiplet architecture, specifically designed for cutting-edge large language model training and inference. REBEL-Quad uses unified mixed-precision cores, executing FP8 and FP16 in a single pipeline without requiring separate compute blocks or kernel recompilation.
Specifications
| Model | Compute | Memory | Interface | TDP | Process |
|---|---|---|---|---|---|
| REBEL-Quad | 2,048 TFLOPS (FP8) / 1,024 (FP16) | 144GB HBM3E | 2× PCIe Gen5 x16 | 600W | 5nm |
Official Website
Driver Downloads
Linux
Related Documentation
OS Support
| Windows | Linux | macOS | Android |
|---|---|---|---|
| ❌ | ✅ | ❌ | ❌ |
Version History
| Version | Release Date | Description |
|---|---|---|
| REBEL-Quad | 2025 | 4-chiplet UCIe-Advanced architecture, 144GB HBM3E |
Performance Benchmarks
| Model | Task | Performance Metric |
|---|---|---|
| REBEL-Quad | FP8 Inference | 2,048 TFLOPS |
| REBEL-Quad | FP16 Training | 1,024 TFLOPS |
Pricing
| Model | Reference Price | Notes |
|---|---|---|
| REBEL-Quad | Contact vendor | Enterprise data center product |
Quick Installation
Linux
# Install Rebellions SDK
# See official documentation
pip install rebellions-sdk
Code Examples
Python (Rebellions SDK)
import rebellions as rb
# Initialize device
device = rb.Device(0)
x = rb.randn((1024, 1024), device=device)
y = rb.matmul(x, x)
print(f"REBEL-Quad matrix multiply: {y.shape}")
Architecture Highlights
- UCIe-Advanced Chiplet: 4 homogeneous chiplets interconnected via UCIe, 1TB/s bidirectional bandwidth, 11ns latency
- Mixed-Precision Pipeline: FP8 and FP16 execute in a single pipeline with 2.8× higher compute density than ATOM™
- Predictive DMA: Software-controlled DMA engine with 2.7TB/s effective bandwidth, reducing long-context LLM latency
- Full-Mesh Synchronization: Hardware-accelerated full-mesh synchronization across 256 routers, maintaining high utilization under sparse or imbalanced loads
Model Compatibility
| Model/Framework | Support | Notes |
|---|---|---|
| PyTorch 2.x | ✅ Native support | vLLM and Triton compatible |
| vLLM | ✅ | Inference optimization |
| Llama Series | ✅ | Large-scale deployment |
| Transformer LLM | ✅ | Cutting-edge model training |
Related Products
If you're evaluating alternatives, the following products may also fit your scenario:
- NVIDIA B200 SXM — NVIDIA (GPU Graphics Processor)
- AMD MI300X — AMD (GPU Graphics Processor)
- Cerebras WSE-3 — Cerebras (ASIC Dedicated Accelerator)
- AWS Trainium / Inferentia — Amazon AWS (ASIC Dedicated Accelerator)
- Intel Gaudi (Habana) — Intel (ASIC Dedicated Accelerator)
- SambaNova RDU — SambaNova (ASIC Dedicated Accelerator)
- Cambricon Siyuan MLU — Cambricon (ASIC Dedicated Accelerator)