GPU & Accelerated Computing Bundle

CUDA and GPU-accelerated data science (RAPIDS/cuDF, cuPyNumeric), HPC and Slurm workflows, distributed LLM training & inference (NeMo, Megatron), numerical optimization (cuOpt), and quantum computing (CUDA-Q) for accelerated-computing engineers.

Browse all skill bundles

Included Skills (2)

  1. Gke AI Troubleshooting Tpu Performance Degradation — The skill diagnoses throughput drops and step-time regressions in GKE Cloud TPU training, determining whether hardware, network fabric, or workload resource bottlenecks cause the issue. It uses ML Diagnostics Workload Monitoring and minute-level Cloud Monitoring metrics to pinpoint TPU duty-cycle reductions of 15% or more. Data scientists and ML engineers benefit by resolving performance degradation without pod crashes or full execution stalls.
  2. Gke AI Troubleshooting Tpu Mxla Hang — The skill diagnoses multi‑slice GKE Cloud TPU training hangs by analyzing Megascale XLA logs and Cloud Monitoring latency metrics. It distinguishes compiler or launch divergence from host data stalls and TPU, SparseCore, ICI, or network faults, enabling engineers to pinpoint the root cause of frozen training jobs that emit HANG_DETECTED. This service is ideal for teams experiencing abrupt training stalls.

Opening your workspace…