Skymizer Launches HyperThought: Build Your Own AI Chip with Skymizer's LPU IP
HyperThought is Skymizer's next-generation LPU IP for real-time, multimodal, and agent-based AI — letting chip designers build their own AI silicon.
Skymizer today launched HyperThought™, its next-generation Language Processing Unit (LPU) IP. Built on an Octa-Core HTX301 processor and the LISA v3 instruction set architecture, HyperThought is purpose-built for LLMs and optimized for performance and power efficiency.
Build your own AI chip
With HyperThought, chip designers and system builders can develop their own AI silicon on a mature T28nm process using standard LPDDR4/5 memory — bringing intelligent, multimodal, agentic AI to the edge and on-prem without cutting-edge fabrication.
- Octa-Core HTX301: 240 tokens/second for Llama2 7B prefill
- Multi-chip scaling: up to 1200 tokens/second for Llama 7B prefill
- Model reach: up to 600B parameters
- Efficient by design: weight compression outperforming llama.cpp by 9%–17.8%, KV-cache compression with minimal perplexity loss
HyperThought marks Skymizer’s transition from AI compiler specialist to chip IP supplier.