AI BriefingSkymizer unveils HTX301
·2026.05.08 10:36
Key point
Skymizer has announced the HTX301 PCIe inference card, featuring 384GB of memory and around 240W power draw.
Details
Skymizer has unveiled HTX301, the first reference chip based on its HyperThought platform.
- The company stated that by combining 6 HTX301 chips with 384GB of memory, it can run 700B-class LLM inference on a single PCIe card using about 240W.
- It targets enterprise on-premise inference without large GPU clusters, NVLink/NVSwitch, or complex cooling infrastructure.
- The core is prefill/decode task separation and decode-first silicon, with software orchestration explained to improve latency and power efficiency.
- Through LISA (Language Instruction Set Architecture), it claims to unify on-device and on-premise deployment for models from 4B to 700B under a single architecture.
Skymizer positioned this as new AI infrastructure for data sovereignty, predictable cost, and deterministic performance.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Hierarchical Memory Architecture Needed to Resolve AI Inference Bottlenecks
2026 AI Inference Hardware Innovations: New Architectures for Prefill/Decode Disaggregation and Solving Memory Bottlenecks