Open Source

Skymizer's HTX301 PCIe card packs 384GB memory at just 240W for AI inference

A low-power inference card with 384GB of memory – could it disrupt NVIDIA’s dominance?

Deep Dive

Key Points
  • 384GB of onboard memory – enough to hold a 200B-parameter model in 16-bit precision without sharding
  • Only ~240W TDP, far below NVIDIA H100 (700W) or B200 (1000W), enabling denser server configurations
  • Built by Taiwan-based Skymizer, leveraging proprietary accelerator IP and compiler software for efficient inference

Why It Matters

Enables cost-effective, memory-rich inference for giant LLMs on a single PCIe card at low power.

📬 Get the top 10 AI stories daily