What is Recogni?
Recogni, now operating as Tensordyne, develops the Napier AI inference system for data centers. The platform uses logarithmic math and custom 3nm silicon to process transformer-based models, targeting hyperscalers, cloud providers, and enterprises. The hardware is scheduled to enter high-volume manufacturing in 2026, with an SDK listed as coming soon.
What are the features of Recogni?
- Logarithmic math architectureUses proprietary log-math (TDN Math) to accelerate transformer-based generative AI inference workloads.
- Custom 3nm inference chipPurpose-built silicon (TDN AIP) taped out with Broadcom and TSMC to handle multi-modal generative AI inference.
- Cell-based scale-up interconnectLow-latency networking (TDN Link) designed for linear scaling across compute trays (TDN ACT).
- Air-cooled rack designOperates at 30 kW per pod without requiring liquid cooling infrastructure.
- Framework integrationSupports PyTorch, Triton, and vLLM for deployment within Kubernetes-managed stacks.
- 16-bit precision inferenceRuns large models in uncompromised 16-bit precision.
What are the use cases of Recogni?
- Real-time video generationProduces 4K video content at 30 frames per second using the cell-based scale-up interconnect.
- Multi-trillion parameter MoE servingOptimized for models like DeepSeek-V4 using EP72 parallelism and TDN Link.
- High-speed agentic codingDelivers over 1,000 tokens per user for low-latency code generation and multi-step reasoning.






