GPU Workload on EKS Cluster

admin

Today I deployed my first GPU accelerated workload using the Karpenter autoscaler and the NVIDIA gpu-operator running on Amazon Web Services (AWS)'s Elastic Kubernetes Service 🙂

$ kubectl logs -l app=pytorch-mnist
Train Epoch: 2 [55680/60000 (93%)]  Loss: 0.161822
Train Epoch: 2 [56320/60000 (94%)]  Loss: 0.008159
Train Epoch: 2 [56960/60000 (95%)]  Loss: 0.217744
Train Epoch: 2 [57600/60000 (96%)]  Loss: 0.042390
Train Epoch: 2 [58240/60000 (97%)]  Loss: 0.055086
Train Epoch: 2 [58880/60000 (98%)]  Loss: 0.090941
Train Epoch: 2 [59520/60000 (99%)]  Loss: 0.156907

Test set: Average loss: 0.0345, Accuracy: 9890/10000 (99%)