FuriosaAI RNGD
Vendor: FuriosaAI (South Korea)
Category: ASIC Dedicated Accelerator
Architecture: TCP (Tensor Contraction Processor)
Introduction
FuriosaAI RNGD is a data center inference chip from South Korean AI chip company FuriosaAI, based on the proprietary Tensor Contraction Processor (TCP) architecture. RNGD is manufactured using TSMC 5nm process with a 180W PCIe design that can be plugged directly into standard servers without special cooling. FuriosaAI was acquired by NVIDIA in December 2025 for approximately $20 billion.
Specifications
| Model | Compute | Memory | Interface | TDP | Process |
|---|---|---|---|---|---|
| RNGD | Inference optimized (FP8/INT8) | HBM | PCIe Gen5 | 180W | TSMC 5nm |
Official Website
Driver Downloads
Linux
Related Documentation
OS Support
| Windows | Linux | macOS | Android |
|---|---|---|---|
| ❌ | ✅ | ❌ | ❌ |
Version History
| Version | Release Date | Description |
|---|---|---|
| RNGD | 2024 | TCP architecture, TSMC 5nm, 180W PCIe |
| 3rd Generation (in development) | 2028 estimated | 2nm, HBM4, Broadcom packaging |
Performance Benchmarks
| Model | Task | Performance Metric |
|---|---|---|
| RNGD | LLM Inference | High efficiency tokens/watt |
| RNGD | Agentic AI | Low-latency inference |
Pricing
| Model | Reference Price | Notes |
|---|---|---|
| RNGD | Contact vendor | Enterprise data center product |
Quick Installation
Linux
# Install FuriosaAI SDK
# See official documentation
pip install furiosa-sdk
Code Examples
Python (FuriosaAI SDK)
import furiosa
# Initialize device
model = furiosa.compile_model("model.onnx", target="rngd")
output = model.run(input_data)
Architecture Highlights
- TCP (Tensor Contraction Processor): Proprietary tensor contraction processor architecture, purpose-built for inference
- 180W PCIe Design: Standard server compatible, no special cooling required
- Deterministic Execution: Compiler-controlled deterministic execution model with predictable latency
- Mature SDK: Inference compiler supporting mainstream AI frameworks
Model Compatibility
| Model/Framework | Support | Notes |
|---|---|---|
| PyTorch | ✅ | Via compiler conversion |
| ONNX | ✅ | Native support |
| Llama Series | ✅ | Inference optimized |
| Transformer LLM | ✅ | Agentic AI workloads |
Related Products
If you're evaluating alternatives, the following products may also fit your scenario:
- Groq LPU v1 — Groq (LPU Language Processor)
- Cerebras WSE-3 — Cerebras (ASIC Dedicated Accelerator)
- Etched Sohu ASIC — Etched (ASIC Dedicated Accelerator)
- NVIDIA GPU / CUDA — NVIDIA (GPU Graphics Processor)
- AWS Trainium / Inferentia — Amazon AWS (ASIC Dedicated Accelerator)
- Intel Gaudi (Habana) — Intel (ASIC Dedicated Accelerator)
- SambaNova RDU — SambaNova (ASIC Dedicated Accelerator)