A tensor is a multidimensional array of numbers. In Triton, kernels operate on tensor tiles—chunks processed by individual program instances—rather than expressing work one scalar thread at a time. The compiler maps tile operations onto GPU hardware. section 1.1.1
A matrix is two-dimensional in the computation but linear in memory. Its shape, strides, tile dimensions, and program ordering determine which values are loaded together and reused on-chip. section 5.0
A Triton reduction operates on the tl.tensor held by one program instance: its axis reduces a tile dimension, not the full tensor. A grid-wide reduction needs additional orchestration, such as a second kernel or atomics. section 4.3
Tensor operations appear in GPU workloads such as:
- Matrix multiplication, where tiled data feeds a tl.dot kernel. section 5.0
- Reductions, including sums, maxima, and custom combinations over a tile. section 4.3









