← Home
CHIPS

Revolutionary AI Accelerator Challenges Nvidia and AMD

May 14, 2026 Priya Nair

Old Tech, New Tricks

Skymizer, a tiny company, unveiled a PCIe AI accelerator on May 10, 2026, capable of running massive language models locally with low power consumption.

The accelerator uses older technology, including decade-old DDR4 memory and 28nm chips, to achieve its remarkable performance while sipping just 240W of power.

Skymizer's innovative design leverages LPDDR memory and older chips to deliver impressive results, potentially embarrassing Nvidia and AMD with their more modern, power-hungry cards. This unusual approach enables the company to challenge industry giants.

Can Legacy Tech Keep Up with AI Demands?

The AI accelerator can run 700 billion large language models (LLMs) locally, a feat that typically requires significant computational resources and energy. Skymizer's achievement is a testament to the company's creative chip design.

Despite using older technology, Skymizer's PCIe AI accelerator demonstrates that innovative design can overcome the limitations of legacy tech. The company's success raises questions about the future of AI accelerator development.

The introduction of Skymizer's AI accelerator may have significant consequences for the industry, potentially forcing Nvidia and AMD to rethink their strategies. As AI continues to grow, companies that can deliver innovative, power-efficient solutions will be well-positioned for success.

Frequently Asked Questions

What makes Skymizer's AI accelerator unique? Skymizer's AI accelerator uses older technology, such as DDR4 memory and 28nm chips, to achieve low power consumption while running massive language models.

How much power does the AI accelerator consume? The Skymizer AI accelerator sips just 240W of power, significantly less than many modern AI accelerators.

Can the AI accelerator run large language models locally? Yes, Skymizer's PCIe AI accelerator can run 700 billion LLMs locally, a remarkable feat that typically requires significant computational resources.

Read full article on Tech Site News →