Tensordyne has unveiled a groundbreaking AI-inference rack, claiming to achieve 13 times the throughput of NVIDIA’s NVL72 GB300 in an internal simulation comparison. This advancement is rooted in Tensordyne’s Napier platform, which employs logarithmic mathematics to convert multiplication operations into addition, thereby significantly reducing energy consumption and chip area dedicated to compute. The TDN rack reportedly reaches a capacity of 363,000 tokens per second, demonstrating a notable efficiency advantage over comparable NVIDIA systems, where Tensordyne asserts it can serve six times more tokens while using 30% less power in the same data center footprint. The TDN72 pod further enhances performance by integrating 72 processors with an any-to-any scale-up interconnect, minimizing data movement latency for complex models.

Tensordyne: Tensordyne is a US-based AI systems company specializing in next-generation inference hardware and platforms. It recently announced its Napier inference system built around proprietary logarithmic mathematics to optimize AI computations at the math level. This development is central to the news, as it underpins claims of major gains in rack throughput and power efficiency for serving models like DeepSeek-R1.
TDN AI Processor: The TDN AI Processor, also known as the Napier chip, is Tensordyne’s custom AI inference processor manufactured on TSMC’s 3nm process node. It integrates substantial on-chip SRAM with HBM3E memory and supports logarithmic math optimizations to reduce compute area and power demands. It forms the core hardware enabling the high-density TDN72 pods and racks highlighted in the announcement.
Nvidia NVL72 GB300: NVIDIA NVL72 GB300 is NVIDIA’s rack-scale AI platform integrating 72 Blackwell Ultra GPUs in a liquid-cooled configuration designed for high-performance inference and training workloads. It serves as the direct benchmark in Tensordyne’s comparisons for token throughput and infrastructure requirements in the news.

`json
{
“Efficiency”: “The TDN inference racks deliver significantly greater token throughput within the same data center footprint while using less power than comparable NVIDIA rack systems.”,
“Technology”: “Tensordyne’s Napier platform employs logarithmic mathematics for AI inference, converting multiplications into additions to minimize energy consumption and transistor usage in compute operations.”,
“Architecture”: “The TDN72 pod features an any-to-any scale-up interconnect across 72 processors, reducing data movement latency in large models such as mixture-of-experts architectures.”
}
`