Which AI chip is stronger? The Olympic Games of the artificial intelligence world announced the answer

Financial Associated Press, April 4 (Editor Huang Junzhi) According to the latest results of MLPerf, the world's authoritative AI benchmark test, Qualcomm Inc's artificial intelligence chip beat Nvidia in two of the three energy efficiency indicators. And a Taiwanese start-up came out on top in one of the remaining metrics.

It is reported that Nvidia occupies a dominant position in the artificial intelligence model training market. But after these AI models are trained, they are widely used in so-called "inference" work, such as generating textual responses to prompts and determining whether an image contains a cat.

Analysts believe that as companies apply artificial intelligence technology to their products, the market for data center inference chips will grow rapidly, but many companies, including Google, have begun to explore how to control the additional costs that will increase.

It is understood that one of the main costs is electricity. Qualcomm used its experience in designing chips for battery-driven devices such as smartphones to develop a chip called CloudAI100, which aims to reduce power consumption.

Qualcomm's AI100 beat Nvidia's flagship chip, the H100, at image classification in test data released by MLCommons on Wednesday, based on how many data center server queries each chip can perform per watt. The Qualcomm chip achieved 197.6 server queries per watt, compared to Nvidia's 108.4 per watt. Taiwanese startup Neuchips topped the list with 227 queries per watt.

MLCommons is a third-party organization that measures the processing speed of artificial intelligence. Its subsidiary MLPerf was initiated by Turing Award winner David Patterson (David Patterson) in conjunction with Google, Stanford, Harvard University and other top academic institutions. It is the most authoritative and influential international AI performance benchmark test, and its status is equivalent to the global The "Olympic Games" in the field of AI.

Also, Qualcomm beat Nvidia in object detection with a score of 3.2 per watt compared to Nvidia's 2.4 per watt. Object detection can be used in applications such as analyzing video footage of a retail store to see where shoppers are most frequented.

Still, Nvidia came out on top in both absolute performance and power efficiency in a test of natural language processing, the most widely used artificial intelligence technology in systems like chatbots. Nvidia achieved 10.8 samples per watt, while Neuchips came in second with 8.9 samples per watt, and Qualcomm was third with 7.5 samples per watt.