Skymizer HTX301 – IP Licensable AI Accelerator Card with 384GB for 700B Models (skymizer.ai)

Skymizer has announced the HTX301, an innovative AI accelerator card designed to handle ultra-large language models (LLMs) from 4 billion to 700 billion parameters without the need for extensive GPU clusters. This new architecture shifts away from traditional superscalar GPU designs by disaggregating prefill and decoding workloads, utilizing a decode-first silicon approach that leverages the LISA™ instruction set. This advancement could significantly reduce the complexity and cost associated with deploying large-scale AI models by eliminating reliance on NVLink/NVSwitch interconnects and elaborate cooling systems. The HTX301's architecture promises data sovereignty and predictable performance, as all inference is conducted on-premises, keeping sensitive data secure. With the ability to scale efficiently across varying model sizes, users can provision AI systems based on actual requirements instead of worst-case scenarios. This means organizations can achieve enterprise-grade privacy, power efficiency, and simplified deployment, marking a potential paradigm shift in how ultra-large AI models are implemented across industries.