DeepSeek Hits China’s Chip Wall: The Tech Hurdle

Can Chinese AI companies handle the deployment of DeepSeek? With hardware limitations and delays compared to Nvidia, it's a significant challenge for China's chip industry.

The Challenge for Chinese AI Companies

The buzz around the adoption of DeepSeek by Chinese companies has taken center stage. “Can Chinese AI chip companies support the deployment of DeepSeek models?” This pertinent question calls for a detailed analysis.

Understanding DeepSeek: Full vs. Distilled Versions

DeepSeek offers two main versions of its models. The distilled version is smaller, with billions or tens of billions of parameters, offering lower deployment costs. On the other hand, the full version of DeepSeek boasts up to 671 billion parameters.

These versions pose different challenges for the Chinese AI industry. While Chinese AI chips can handle the distilled version of DeepSeek, supporting the full version is a different story.

The Struggle of Chinese Chips to Support the Full Version of DeepSeek

Running the full version of DeepSeek requires more than just an 8-card server rack, presenting a challenge for Chinese chip companies. Jack, a senior AI chip engineer, points out: “It will be tough, and there may be no end in sight for successfully adapting to the full version of DeepSeek.”

Nvidia vs. Chinese AI Companies: A Technological Lag?

Nvidia, a recognized market leader, holds a significant edge in this competition. Chinese AI chips do not support FP8 data representations, meaning that deploying DeepSeek requires twice as much storage space and memory, necessitating more cards.

Bo Lin, a veteran in the chip industry, candidly states that the latest Chinese AI chips do not support FP8: “This shows that many people making AI chips in China don’t understand AI.” Although Chinese companies are setting data processing speed records, these advances are not enough to compete with Nvidia. It seems there is still a long journey ahead for Chinese AI firms.