Neuchips Viper Series LLM Inference Card

Neuchips Corporation is a Taiwanese AI ASIC designer working since 2019, headquartered in Zhubei, Hsinchu County, with offices in Taipei and Silicon Valley. After building the RecAccel recommendation accelerator, it pivoted to generative AI, aiming to let enterprises run large language models offline at low power instead of relying on cloud GPUs.

The Viper Series is a half-height, half-length, single-slot PCIe card built on the 7nm Raptor N3000 LLM inference ASIC. It carries 64GB of LPDDR5 memory, runs at 45W by default, and supports models up to 14B parameters on a single card, including Llama, Mistral, Gemma, Phi, Qwen, TAIDE and DeepSeek distilled models. An SDK with a compiler, PyTorch extension and FFP8 quantization supports on-premises RAG, document analysis, customer service and knowledge management. Partners such as Vecow, ShareGuru and Netiotek build workstations and enterprise solutions on it.

• Raptor N3000 ASIC: 7nm LLM inference chip designed for low-power generative AI.
• Compact PCIe card: half-height, half-length, single-slot design with active or passive cooling.
• Large memory: 64GB LPDDR5 at 6400 Mbps for models and vector database storage.
• Configurable power: 25W to 75W board power with 45W default.
• Broad model support: Llama, Mistral, Gemma, Phi, Qwen, TAIDE and DeepSeek distilled models up to 14B parameters.
• Software SDK: compiler, PyTorch extension, FFP8 quantization, graph optimization and runtime.

A Vietnamese bank or insurer can run an internal document question-answering assistant on Viper-equipped servers without sending data to the cloud. System integrators can build offline RAG workstations for law firms and government agencies. Manufacturers can deploy knowledge management assistants on factory networks isolated from the internet.

A single 45W card runs LLMs of up to 14B parameters, so generative AI can be added to standard servers and workstations without high-power GPUs. Data stays on-premises for privacy and data sovereignty. The company states the chip offloads over 90% of generative AI resource needs from the CPU. The low-profile form factor fits compact and edge systems.

Enterprises and public bodies increasingly want private, on-premises generative AI for sensitive documents, while GPU costs and power limits constrain adoption. Low-power inference accelerators for local LLMs are a growing category. In Vietnam and Southeast Asia, data localisation rules and limited data-centre power favour compact offline AI appliances.

Working since 2019; headquartered in Zhubei, Hsinchu County, Taiwan, with offices in Taipei and Silicon Valley. Chairman, founder and CEO Dr. Youn-Long Lin. Raised Series A, Series B and a USD 17.6 million Series B3 round; investors shown include JAFCO Asia, Powerchip, Raydium, Sunplus and Wistron. Founding member of MLCommons. Partners include Vecow, ShareGuru and Netiotek.