Skymizer HTX301 – IP Licensable AI Accelerator Card with 384GB for 700B Models

A single PCIe card designed for on-prem inference of ultra-large language models, HTX301 packs six LPU chips and 384 GB of on-card memory to deliver up to 700 billion-parameter model inference at about 240 W total power. Built on a mature 28 nm process with an LPDDR-class memory subsystem, it uses the LISA instruction set and a unified architecture that scales from 4B to 700B parameters without over-provisioning. The card emphasizes a compact deployment footprint for enterprise and agentic AI workloads, aiming to avoid multi-node GPU clusters and their NVLink/NVSwitch interconnects, orchestration complexity, and heavy cooling requirements.

The design disaggregates prefill and decode workloads and adopts a decode-first silicon approach to optimize throughput for generative inference, trading the superscalar GPU-cluster model for deterministic, predictable performance on a single card. That approach is positioned to preserve data sovereignty by keeping weights and prompts on-premises, reduce capital and operational costs tied to cluster infrastructure, and simplify scaling by provisioning for the model in use rather than worst-case needs. Early-access evaluation hardware and registration are available for enterprises, researchers, and integrators interested in testing the platform.