cuda core

CUDA cores are NVIDIA GPU's fundamental processing units, each capable of executing floating-point and integer operations, with thousands per GPU enabling massive parallel computation for AI and graphics workloads. CUDA core function: each core executes one thread's operations; arithmetic logic unit (ALU) plus floating-point unit (FPU). Core count: ranges from hundreds (laptop GPUs) to 10,000+ (datacenter GPUs like H100); more cores = more parallel capacity. Streaming Multiprocessor (SM): cores grouped into SMs (typically 64-128 cores per SM); SM has shared resources (memory, schedulers). Parallel execution: SM schedules warps (32 threads) onto cores; cores execute in lockstep. Clock speed: typically 1-2 GHz; raw performance = core count × clock × operations per clock. Beyond CUDA cores: Tensor Cores (matrix operations), RT Cores (ray tracing), and specialized units increasingly important. Memory bandwidth: cores must be fed data; memory bandwidth often the bottleneck, not core count. Occupancy: percentage of SM resources used; more warps can hide memory latency. CUDA programming: write kernels that launch across cores; CUDA Runtime handles scheduling. Comparison to CPU cores: CPU cores complex (out-of-order, branch prediction); GPU cores simple but numerous. Performance: A100 has 6,912 CUDA cores + 432 Tensor Cores; H100 has 16,896 CUDA cores + 528 Tensor Cores. CUDA cores provide the parallel horsepower for GPU computing.