Skymizer's HTX301 Challenges Nvidia and AMD with Efficient AI Accelerator Using Older 28nm Technology
Key Takeaways
- ▸HTX301 consumes 240W—less than half the power of Nvidia's RTX PRO 6000 Blackwell (600W) and comparable AMD solutions—while supporting models with up to 700 billion parameters
- ▸Uses older 28nm process technology with LPDDR4/5 memory, proving that algorithmic efficiency and memory architecture can compete with cutting-edge silicon
- ▸Designed for on-premises deployment in standard servers, eliminating expensive infrastructure redesigns and addressing data sovereignty and cost unpredictability
Summary
Skymizer, a Taiwanese AI hardware startup, has unveiled the HTX301, a PCIe AI accelerator card that challenges Nvidia and AMD's dominance by using an unconventional approach: older 28-nanometer chips paired with standard LPDDR4/5 memory instead of modern HBM technology. The card can run language models with up to 700 billion parameters while consuming only 240 watts—less than half the power required by comparable cards from Nvidia and AMD—and offers 384GB of total memory capacity. Built on Skymizer's HyperThought platform, the HTX301 delivers 30 tokens per second and uses compression techniques for weights and KV cache.
The accelerator is specifically designed for on-premises deployment, addressing enterprise concerns about data sovereignty, privacy risks, and unpredictable cloud costs. Unlike competitors that require expensive data center infrastructure redesigns, the HTX301 fits into standard air-cooled servers without modification. Skymizer will preview the card at Computex 2026, where independent verification will test whether the impressive specifications translate to real-world performance. The company claims the HTX301 signals the end of the hyperscale GPU cluster era for large language model inference.
- Real-world testing at Computex will be critical; the company joins a history of AI hardware startups that have struggled to deliver on ambitious claims
Editorial Opinion
If Skymizer's performance claims hold up under real-world testing, the HTX301 could genuinely disrupt the AI accelerator market by proving that algorithmic efficiency matters as much as cutting-edge silicon. The company's counterintuitive approach—using decades-old process technology to compete with Nvidia and AMD—suggests a refreshing focus on practical efficiency over raw processing power, potentially democratizing on-premises AI infrastructure for cost-conscious enterprises.
