Tensordyne Napier chip uses logarithmic math to challenge Nvidia in AI inference. By IT News / Tue, Jun 16 2026 / Tensordyne has announced the "Napier" AI accelerator, a specialized chip designed to optimize the inference process for large-scale AI models. The hardware utilizes logarithmic mathematics to convert complex multiplications into simpler additions, significantly reducing the computational overhead required for model execution. This architectural approach allows for smaller processing units, creating space for integrated SRAM alongside 144 GB of high-bandwidth HBM3E memory. A full Tensordyne Napier Rack is engineered to deliver 608 PFlops of computing power while maintaining a power consumption of 120 kilowatts under full load. The company claims this configuration can outperform current Nvidia GB300 technology by delivering up to 13 times more tokens per second depending on the specific AI model. Furthermore, the system is designed for high efficiency, reportedly offering up to 17 times better performance per watt than competing solutions. The startup has already secured over $200 million in orders for these systems, which are compatible with popular models like Llama 3.1 and DeepSeek-R1. Developed in collaboration with Broadcom and manufactured by TSMC, the Napier chip supports various data formats including FP16, FP8, and Int8. While Tensordyne prepares for delivery, the company faces upcoming competition from Nvidia's Groq 3 LPX systems, which are also specifically optimized for AI inferencing tasks.

Tensordyne tapes out new inferencing chip. Company aims to take on Nvidia with new chip June 16, 2026 AI chip firm Tensordyne has started production of its new inferencing chip. Developed in partnership with HPE Juniper Networks and Broadcom, Tensordyne has successfully completed tape-out of its Tensordyne Napier artificial intelligence processor (TDN AIP), now being manufactured by TSMC on its 3nm process node. "The market is hungry for fast AI; customers want speed, but achieving it has always meant accepting prohibitive costs," said Marc Bolitho, Tensordyne CEO. "By optimizing math, compute, memory, and networking from first principles, Napier delivers affordable inference without compromising on speed. We're delivering fast, cost-effective AI and eliminating the need for bolted-together systems that require multiple racks." Tensordyne said it has secured more than a dozen Letters of Intent from hyperscaler, neocloud, and sovereign AI infrastructure operators, representing more than $200 million in value. The company is set to place its TDN AIP's inside a TDN 72 - a 72-chip inference pod that it claims will outperform a full Nvidia NLV72 rack on a tokens per watt basis, while requiring less energy. Each air-cooled pod offers 76.7 petaflops of FP16 compute power and requires 30kW. A full Tensordyne Napier AI Inference rack would comprise four TDN 72 pods, suggesting 306 petaflops of FP16 per rack and 120kW per rack. Frank Ostojic, SVP and general manager, ASIC Products Division at Broadcom, says, "We are happy to support the Tensordyne team as our partner & customer. They are leveraging Broadcom's fundamental IP Platform, advanced Silicon and Packaging design technologies to bring their 3nm AI Inference Accelerator to market - focused on achieving industry-leading performance and power efficiencies." The company is set to raise new funding via an anticipated Series D later this year. It has raised more than $211 million to date, with investors including Plug and Play, Mayfield Fund, Bosch Ventures, Toyota Ventures, GreatPoint Ventures, and others. It raised $102m in Series C funding in 2024. "Cirrascale is looking forward to introducing the Tensordyne Napier system to our customers to unlock a new level of energy-efficient, high-performance AI infrastructure at scale. What Tensordyne is building directly addresses the pent-up demand for alternative technologies that provide fast, cost-efficient inference," adds Dave Driggers, CEO of Cirrascale Cloud Services, a Tensordyne customer. Formerly Recogni, Tensordyne had been designing chips for autonomous vehicles since it was founded in 2017. The company has since pivoted to develop AI and inference chips instead. More in IT hardware & semiconductors.