New PCIe AI accelerator promises massive enterprise language models - TechRadar

Taiwanese company Skymizer has unveiled the HTX301 PCIe AI accelerator, which uses 28nm legacy process technology to run large language models with up to 70 billion parameters while consuming only 240... Read original →

Skymizer’s HTX301 leverages legacy 28nm nodes to run 70B-parameter LLMs at just 240W—a direct challenge to the industry’s obsession with bleeding-edge nodes and HBM. By substituting LPDDR4/5 for costly HBM/GDDR, it slashes BOM and sidesteps export-controlled memory supply chains, appealing to enterprises prioritizing data sovereignty under tightening U.S.-China tech restrictions. This forces NVIDIA and AMD to accelerate low-power inference SKUs, especially for non-U.S. markets. However, 28nm’s poor area efficiency risks yield bottlenecks if foundry capacity isn’t secured. If validated in production over the next 18 months, this approach will catalyze a wave of chiplet-based, compression-optimized LPUs, fragmenting the AI accelerator market beyond GPU hegemony.