CUDA Hello World: Done Less Wrong (ashvardanian.com)

A common CUDA “Hello World” pattern—launching kernels with NVCC’s triple-bracket <<<>>> syntax—works interactively but is brittle in production. The author shows that those kernel launches return void and hide submission-time failures on complex systems (multi-GPU HGX/DGX nodes with asynchronous CPU activity). They recommend treating launches as high-latency, asynchronous operations: explicitly order work with streams, always check CUDA API errors, and avoid relying solely on <<<>>> for correctness. Technically, the post demonstrates safer alternatives: create and use cudaStream_t for ordering and async launches (kernel<<<grid,block,shared,stream>>>), but for real error reporting use the CUDA Driver/Execution Control API via cudaLaunchKernel which returns submission errors (e.g., an intentional 1GB shared-memory request produces "invalid argument" at launch). It shows how to pass arguments with a void* kernel_args[] and uses cudaMallocManaged to simplify host/device memory. Finally, it cautions that higher-level synchronization (cooperative_groups::sync for whole-grid sync) exists but has constraints and must be used carefully. The takeaway: adopt streams, explicit error checks, and driver-level launches to avoid silent failures and improve robustness on modern heterogeneous GPU systems.