RNGD
150W AI inference accelerator chip by FuriosaAI
FuriosaAI designed the RNGD, an artificial intelligence accelerator focused on efficient inference. Built on TSMC’s 5 nm process, it uses tensor contraction to deliver high performance per watt for large language models in data centres.
- Built on TSMC 5 nm process with 150 W TDP
- Uses tensor contraction instead of matrix multiplication
- Integrates 48 GB of HBM3 memory via 2.5D packaging
- Delivers high efficiency for large language model inference
- Operates at 1 GHz to maximize power efficiency
What it is
The RNGD, pronounced Renegade, is an artificial intelligence accelerator chip designed by FuriosaAI. It targets inference workloads, particularly for large language models, within standard data centre infrastructure. The chip aims to provide high efficiency, allowing operators to add AI capacity without increasing power limits significantly. It is implemented on a PCIe card and integrates natively into existing serving pipelines and Kubernetes clusters.
How it came about
FuriosaAI presented the RNGD at Hot Chips 2024. The company achieved its performance targets just three months after receiving silicon from the foundry. The design focuses on flexibility to allow for future software optimizations. LG AI Research adopted the technology to scale its EXAONE foundation models, citing reduced total cost of ownership and straightforward integration.
Inside the chip
The architecture uses tensor contraction as its primary operation, which is more general than the matrix multiplication used in many GPUs. This approach allows better handling of varying batch sizes common in inference. The chip consists of Processing Elements that can operate independently or fuse into larger groups. Each element contains a CPU core for control and a Tensor Unit with multiple slices. These slices use substantial SRAM capacity to maximize data reuse and minimize external memory access.
How it was made
TSMC manufactures the RNGD using its 5 nm process node. The chip operates at a conservative 1 GHz clock speed to maintain low power consumption. It employs 2.5D packaging, integrating the system on a chip with twelve-layer HBM3 memory modules on a single interposer. This configuration provides 48 GB of high bandwidth memory while keeping the thermal design power at 150 watts.
Where it went
Data centre operators use the RNGD to expand AI cloud capacity within fixed power envelopes. It supports interactive serving for models like Qwen3-32B and EXAONE. The chip integrates with standard air cooling systems, avoiding the need for liquid cooling infrastructure. Enterprises deploy it for real-time agents, search summarization, and retrieval augmented generation tasks behind corporate firewalls.
Versions and family
The RNGD is available as a PCIe card. FuriosaAI also offers the NXT RNGD Server, an enterprise appliance consuming 3 kW. This server allows bare-metal access for validating full software stacks. The company provides an OpenAI compatible API endpoint for application testing without local installation. Software development kits are updated regularly to improve token output per kilowatt.
Why it matters
The chip demonstrates that specialized hardware can outperform general purpose graphics processors in specific inference tasks. It claims up to twice the efficiency of competing solutions in certain benchmarks. By focusing on power efficiency, it enables regional cloud providers to maximize compute density per megawatt. This approach helps diversify compute catalogs beyond traditional GPU allocations.
Problems and reception
LG AI Research praised the combination of real world performance and cost reduction. Benchmarks compare it favourably against the NVIDIA RTX PRO 6000 and L40S in terms of tokens per second per kilowatt. The company notes that performance is subject to change with software updates. Early adopters highlight the ease of integration into existing virtualized infrastructure.
Questions people ask about RNGD
What makes the RNGD different from a GPU?
The RNGD uses tensor contraction as its core operation rather than fixed matrix multiplication. This allows more flexibility for varying batch sizes in inference. It also prioritizes SRAM capacity over bandwidth to keep data on-chip, reducing power consumption compared to typical graphics processors.
How much power does the RNGD consume?
The chip has a thermal design power of 150 watts. This low power draw allows it to fit into standard air cooled data centre racks without requiring additional power infrastructure. It aims to deliver higher token output per kilowatt than many competing accelerators.
What memory does the RNGD use?
It uses 48 GB of HBM3 memory. The memory modules are integrated with the processor on a single interposer using 2.5D packaging. This setup provides high bandwidth and capacity while maintaining energy efficiency for large model workloads.
Who designed and manufactured the RNGD?
FuriosaAI designed the RNGD architecture and software stack. TSMC manufactured the physical chips using their 5 nm process node. The final product is assembled into PCIe cards for deployment in servers and data centres.
Good reading and sources
There is no English Wikipedia article on RNGD yet, so the facts on this page come from the sources listed above: furiosa.ai/rngd, chipsandcheese.com/p/furiosaais-rngd-at-hot-chips-2024-accelerating-ai. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryChip wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.