Nvidia dominates the market for training AI models with huge amounts of data. But after those AI models are trained, they are put to wider use in what is called “inference” by doing tasks like generating text responses to prompts and deciding whether an image contains a cat.
Analysts believe that the market for data center inference chips will grow quickly as businesses put AI technologies into their products, but companies such as Alphabet Inc’s Google are already exploring how to keep the lid on the extra costs that doing so will add.
One of those major costs is electricity, and Qualcomm has used its history designing chips for battery-powered devices such as smartphones to create a chip called the Cloud AI 100 that aims for parsimonious power consumption.
In testing data published on Wednesday by MLCommons, an engineering consortium that maintains testing benchmarks widely used in the AI chip industry, Qualcomm’s AI 100 beat Nvidia’s flagship H100 chip at classifying images, based on how many data center server queries each chip can carry out per watt.
Qualcomm’s chips hit 197.6 server queries per watt versus 108.4 queries per watt for Nvidia. Neuchips, a startup founded by veteran Taiwanese chip academic Youn-Long Lin, took the top spot with 227 queries per watt.
Discover the stories of your interest
Qualcomm also beat Nvidia at object detection with a score of 3.2 queries per watt versus Nvidia’s 2.4 queries per watt. Object detection can be used in applications like analyzing footage from retail stores to see where shoppers go most often. Nvidia, however, took the top spot in both absolute performance terms and power efficiency terms in a test of natural language processing, which is the AI technology most widely used in systems like chatbots. Nvidia hit 10.8 samples per watt, while Neuchips ranked second at 8.9 samples per watt and Qualcomm was in third place at 7.5 samples per watt.
