Case Study: Orange achieves linear AI/ML scaling with Liqid Matrix CDI

Unleashing AI/ML Superpowers Orange Silicon Valley Transforms Standard Servers with Liqid Matrix CDI

Orange Silicon Valley needed a better way to handle rapidly growing AI and machine learning workloads that demand large amounts of GPU power. Traditional server chassis limit GPU density, leave resources underused, and make it costly and inefficient to scale for tasks like ImageNet training and NLP. To address this, Orange Silicon Valley worked with Liqid and its Matrix Composable Disaggregated Infrastructure (CDI) platform.

Liqid Matrix let Orange Silicon Valley compose 16 NVIDIA A100 GPUs and NVMe storage into a single Dell PowerEdge R6525, enabling direct GPU-to-GPU and GPU-to-storage communication. The result was near-linear scaling from 1 to 16 GPUs, with ImageNet throughput rising to 32,607 images/sec and transformer training reaching 802,600 tokens/sec; with a larger batch size, throughput climbed to 935,343 tokens/sec and minimal validation loss was achieved in under 1 hour 49 minutes.

Orange