GPU Glossary (modal.com)

This announcement is a compact GPU glossary and README-style index that catalogues the key hardware, software, and performance concepts relevant to modern NVIDIA-style GPUs and CUDA workflows. It groups terms into clear sections—Device Hardware (SM, CUDA cores, Tensor Cores, TMA, caches, registers, GPU RAM), Device Software (PTX/SASS, threads/warps/blocks, memory hierarchy), Host Software (CUDA drivers, nvcc, cuBLAS, cuDNN, CUPTI, Nsight, nvidia-smi) and Performance (roofline model, arithmetic intensity, occupancy, warp divergence, memory coalescing, bank conflicts, register pressure, etc.). The format reads like a navigable table of contents for deeper documentation. For the AI/ML community this is significant because GPUs and CUDA concepts are foundational to model training and inference performance. The glossary highlights the building blocks (tensor units, warp scheduling, memory/cache layers) and the profiling/tuning tools and metrics that practitioners use to diagnose bottlenecks—whether compute-bound vs memory-bound, how to improve arithmetic intensity, or how to increase SM utilization and hide latency. As a concise reference, it helps engineers understand trade-offs when optimizing kernels, choosing hardware, or interpreting profiler output, making it easier to squeeze more throughput out of GPUs and to communicate optimization strategies across teams.