Demeter

Project Demeter was a collaborative initiative between Code Zero, Merge IT, and Zadara, designed to benchmark machine learning workloads across various hardware platforms. Using Kaggle datasets covering medical image scanning, genetic biomarker detection, and stock market sentiment prediction, the project compared performance on an AWS instance (NVIDIA T4 GPU) against Zadara zCompute virtual machines (NVIDIA A2 and A40 GPUs). Zadara's HCI platform, powered by 100GB Infiniband Mellanox switching and unified Block, File, and Object storage, served as the primary test environment. Monitoring was handled via CloudWatch on AWS and Prometheus on Zadara. The project demonstrated that intense ML workloads can be executed efficiently on modest hardware, achieving a 32% reduction in processing time for medical image analysis and handling 10,000+ medical scans per day. A key finding was that optimising the software stack has a greater impact than raw GPU power, as GPU memory utilisation remained low throughout testing while most of the load was on disk.