# Skymizer

> Skymizer is a software-hardware co-design IP company (founded 2013, Taipei, Taiwan) that builds LPU (Language Processing Unit) IP and silicon for on-prem and on-device large language model (LLM) inference.

Skymizer licenses LPU IP and integrated software to semiconductor companies, device manufacturers, and system integrators. Built on LISA (Language Instruction Set Architecture) and a decade of AI compiler and runtime expertise, its technology runs large language models efficiently on mature process nodes — using architecture and compression rather than leading-edge silicon.

## Products
- [HTX301](https://skymizer.ai/products/htx301/): Reference chip on the HyperThought platform. Six HTX301 chips with 384GB of memory on a single PCIe card run 700B-parameter LLM inference on-prem at approximately 240W — no GPU cluster, NVLink/NVSwitch, or complex cooling. Scales 4B–700B parameters.
- [HyperThought](https://skymizer.ai/products/hyperthought/): Octa-core, multi-chip-scalable LPU IP on LISA v3 for LLMs up to 600B parameters, on-prem and at the edge. Won "Best IP/Processor of the Year" and "Most Promising Product" at EE Awards Asia 2025.
- [EdgeThought](https://skymizer.ai/products/edgethought/): Single-core, compiler-centric LPU IP for on-device LLM inference on resource-constrained edge devices (IoT, AI PCs, edge servers). Integrates with HuggingFace Transformers, NVIDIA Triton, the OpenAI API, and LangChain.

## Company
- [About Skymizer](https://skymizer.ai/about/): Founding (2013), mission, leadership, and board.
- [Announcements](https://skymizer.ai/announcements/): Product launches, partnerships, and awards.
- [News](https://skymizer.ai/news/): Press coverage and media resources.

## Announcements (most recent first)
- [Skymizer Announces HTX301 — Reinventing On-Prem AI Inference](https://skymizer.ai/announcements/htx301-on-prem-ai-inference/) (2026-04-23) — HTX301 enables 700B-parameter LLM inference locally at just ~240W on a single PCIe card — no GPU cluster required.
- [Skymizer’s HyperThought™ Wins “Best IP/Processor of the Year” and “Most Promising Product” at EE Awards Asia 2025](https://skymizer.ai/announcements/hyperthought-ee-awards-asia-2025/) (2025-12-07) — Skymizer's HyperThought LLM Accelerator IP receives dual recognition at EE Awards Asia 2025.
- [Skymizer and JFE Shoji Electronics Join Forces to Advance AI Acceleration Solutions in Japan](https://skymizer.ai/announcements/jfe-shoji-partnership-japan/) (2025-09-10) — JFE Shoji Electronics becomes an official distributor of Skymizer's AI acceleration solutions in Japan.
- [Skymizer Launches HyperThought: Build Your Own AI Chip with Skymizer’s LPU IP](https://skymizer.ai/announcements/hyperthought-launch/) (2025-05-09) — HyperThought is Skymizer's next-generation LPU IP for real-time, multimodal, and agent-based AI — letting chip designers build their own AI silicon.
- [Inventec, VicOne, and Skymizer to Showcase GenAI Cybersecurity Solution for Smart Cockpits at Tokyo Automotive World](https://skymizer.ai/announcements/inventec-vicone-smart-cockpit/) (2025-01-23) — A three-way collaboration delivering an in-car GenAI security solution that protects against data leaks in smart vehicles.
- [Skymizer unveils Skygenie Auto Concept, Powered by ET2](https://skymizer.ai/announcements/skygenie-auto-concept/) (2024-06-05) — A comprehensive software-hardware integration for automotive, bringing advanced language understanding into the vehicle — powered by EdgeThought (ET2).
- [Skymizer Launches Groundbreaking LLM Accelerator IP for on-device LLM Inferencing, EdgeThought, the game-changer in on-device GenAI era.](https://skymizer.ai/announcements/edgethought-launch/) (2024-06-01) — EdgeThought is a compiler-centric LPU IP engineered to accelerate large language models on resource-constrained edge devices — the game-changer for the on-device GenAI era.

## Key facts
- Founded: 2013, Taipei, Taiwan (office also in Hsinchu)
- Business model: LPU IP + integrated-software licensing
- Core technology: LISA instruction set architecture; decode-first silicon; prefill/decode disaggregation; weight and KV-cache compression
- Process node: mature T28nm
- Flagship figure: 700B-parameter LLM inference on a single PCIe card at ~240W (HTX301)