Artificial intelligence (AI) for image recognition and classification is growing fast, with one study estimating a 2024 market share of USD 2.55 billion, with expectations of growth to USD 4.44 billion over the next five years.(1) With a wide range of use cases that includes everything from expediting medical diagnoses to enabling quality assurance for manufacturing and making online shopping suggestions, using AI to classify images can help organisations analyse visual data quickly to deliver the timely answers they need. For businesses seeking to improve key processes with AI-based image classification, one of the first steps is selecting server hardware that can adequately handle this computationally demanding work.
To determine its suitability for running AI inference workloads such as image classification, Principled Technologies tested a Stratus ztC Endurance 7100 server using a ResNet-50 image classification workload at various levels of precision. Across all three precision levels, we found that the Stratus ztC Endurance 7100 offered strong throughput and low latency for CPU-based inference, showing that it’s a viable platform for AI image classification for various use cases – from those that prioritise accuracy to those that prioritise speed.
Running AI inference workloads on the Stratus ztC Endurance 7100
There are countless critical real-world applications for classifying images using machine learning. AI image classification can help speed up quality assurance in industrial manufacturing by providing an automated method to prevent subpar products from making it to market. Accelerating this portion of the manufacturing process can ultimately get products on the shelf faster to meet customer demands. Quick inference for image classification can also lead to quick suggestions for customers shopping at online retailers, recommending other products based on the items they’ve shown interest in. No matter your specific image classification use case, choosing servers with strong throughput can help you get answers from your datasets more quickly.
ResNet-50 is a convolutional neural network that runs 50 layers deep to quickly perform image classification. Using ResNet-50 models from the TensorFlow framework as well as Intel Reference Models, we ran image classification performance tests at three different precision levels: FP32, bfloat16 and INT8. The benchmark reports throughput in the number of images per second that the system could classify, as well as latency (wait times) during analysis.
Please download the test report below to read on.
Share