Cloud GPUs

Accelerated compute on Google Cloud - what CPUs, GPUs, and TPUs each do well, which one fits a workload, and how to provision and optimize them.

• Accelerated compute - how CPUs, GPUs, and TPUs work, and why the architecture (von Neumann vs. parallel ALUs vs. systolic array) decides which is fastest.
• GPU options - the GPU machine families (A4, A3, A2, G2, N1), how clusters are wired (cluster → block → subblock → node → GPU), and which family fits which workload.
• Provisioning options - the five platforms for provisioning GPU clusters (Cluster Director, Compute Engine, GKE, Cloud Run, Agent Platform) on a control-to-managed spectrum, plus a GKE serving demo.
• Accelerating frameworks - how PyTorch and JAX use CUDA kernels, the XLA compiler, and the PyTorch/XLA and JAX/XLA integrations to run fast on one accelerator and portably across GPU, TPU, and CPU.
• Optimizing GPU usage - ML Productivity Goodput (scheduling, runtime, program), and the decision trees for picking a GPU and tuning it for the best performance per dollar.
• Cheat sheet - one-page course review guide: accelerator types, GPU decision factors, options by workload, provisioning spectrum, and frameworks.