Tensor Core
Definition pending
We've cataloged "Tensor Core" but haven't written a full definition yet. Definitions are hand-curated rather than auto-generated, so it takes time to cover every term.
Want this one prioritized? Email us and we'll bump it.
Practical example
Tensor Cores are specialized GPU units that multiply 4×4 matrices in one clock cycle. They're what make FP16/BF16 training 3–8× faster than FP32 on the same GPU. Every modern NVIDIA GPU (Volta+ for datacenter, Turing+ for consumer) has them. The key number: an H100 has 528 Tensor Cores delivering 990 TFLOPS FP16. Without tensor cores, same GPU = ~60 TFLOPS FP32.
Workflow example
Tensor Cores are used automatically when you run FP16 or BF16 operations in PyTorch. Check utilization: nvidia-smi dmon -s pucv shows SM and Tensor Core activity. If running FP16 inference and tensor core utilization is <10%, your framework is not using tensor cores — common with llama.cpp before CUDA graph support. vLLM and TensorRT-LLM use tensor cores more efficiently than llama.cpp for high-throughput serving.