Run 700B-parameter
models on one card.

Skymizer is a software-hardware co-design IP company founded in 2013 in Taiwan. It builds LPU (Language Processing Unit) IP and silicon for on-prem and on-device large language model inference, drawing on a decade of AI compiler and runtime expertise.

An LPU is a processor purpose-built to run large language models efficiently. Skymizer’s LPU IP is based on its LISA (Language Instruction Set Architecture) and maps LLMs onto efficient silicon on mature process nodes, rather than relying on cutting-edge fabrication.

HTX301 is Skymizer’s 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 — without a GPU cluster, NVLink/NVSwitch, or complex cooling.

A single HTX301 PCIe card scales from 4B to 700B parameters and runs 700B-parameter inference at roughly 240 watts, using 384GB of on-card memory.

HyperThought is octa-core, multi-chip-scalable LPU IP (LISA v3) for LLMs up to 600B parameters, on-prem and at the edge. EdgeThought is a single-core, compiler-centric LPU IP for on-device inference on resource-constrained edge devices. HTX301 is the reference chip that productizes HyperThought.

Skymizer’s core business is licensing LPU IP and integrated software to semiconductor companies, device manufacturers, and system integrators. HTX301 is a reference chip built on that IP.

Skymizer achieves efficiency through architecture and compression — weight compression and KV-cache compression plus a decode-first, prefill/decode-disaggregated design — rather than leading-edge silicon. This keeps cost and power low while running very large models.