Deployment & Scaling · Practitioner
CUDA-X Libraries and Nsight: cuDNN, Compatibility, and Finding the Real Bottleneck
Between CUDA and your framework sits a library layer — cuDNN, cuBLAS, NCCL — plus the profiling tools that tell you which of them is costing you. Learn version compatibility, then learn to read a Nsight timeline instead of guessing.
- 1.5 hours lesson
- intermediate level
- NVIDIA AI Track
- Narrated presentation
What you’ll learn
- cuDNN, cuBLAS, NCCL as the tuned primitive layer frameworks call
- Driver / toolkit / framework version compatibility and the published support matrix
- NGC and official framework containers as the way to avoid assembling CUDA by hand
- Nsight Systems for whole-application timelines; Nsight Compute for single-kernel analysis
- torch.profiler and NVTX ranges for framework-level visibility
- Measure-change-measure discipline instead of speculative optimisation
Inside this lesson
A narrated, slide-by-slide presentation with 1 prediction exercise, 1 knowledge check, 5 videos. You answer as you go, and the platform checks your work.