Why hardware security underlies AI progress

Find the Latest SoC Solutions
for... Automotive... IoT... Security... Audio... Video

Semiconductors in AI have been getting a lot more attention recently, serving as the essential building blocks that power and optimize the intricate algorithms driving AI technologies to unprecedented levels of efficiency and performance.

www.embedded.com/, Jun. 04, 2024 – 

Demand for NVIDIA's AI chips was a story all year in 2023, as the company's stock price surged 234%. Supply chain shortages left companies like Microsoft listing access to GPUs as a risk factor for investors and concerns about intellectual property (IP) are leading to enhanced export controls. The U.S. government carved out specific provisions to address semiconductor security in the AI-focused National Defense Authorization Act for Fiscal Year 2024. In short, AI chips have become big news as businesses increasingly build AI into their products and systems to help boost productivity and profits.

That increased attention, especially around the supply chain, is why chip security is a growing concern. Any company developing AI models needs them to be secure, especially as more business depends on the insights, efficiencies, and work they provide – and as these models grow in complexity and incorporate real-time data to make their own decisions. Security and reliability go hand in hand, and any adversarial attacks could cause unexpected and unpredictable behavior from AI.

Accordingly, these AI models are only as secure as the hardware they run on, making the integrity of the semiconductor chip critical to an AI company's success. The more chips we produce (especially if they're being made in-house), and the more complex the chips become, the more security issues we're going to see. The security of those chips must be ironclad so models behave as expected and end users are protected from harm.

The problem is that fully understanding the security of those chips is incredibly difficult and time consuming. There's often a lack of transparency around the third-party IP used to develop AI chips, and the "black box" nature of AI chips makes vulnerabilities difficult to detect.

click here to read more...

 Back