Results from . Each benchmark has equal weight, divided among its selected datasets.

At 1×, a framework matches the fastest runtime for a problem. At 2×, it takes at most twice as long. Failed measurements remain in the suite but do not contribute successes. Include tags match any selection; exclude tags take precedence.

Workload Tags

Each benchmark/dataset is tagged with workload descriptors, discovered programmatically through tracing. The table below lists each tag and a short description.

Tag	Description
`high-dimensional`	5 or more dimensions in a tensor.
`tensor`	3 or more dimensions in a tensor.
`large-query`	5 or more operands on one line.
`elementary-ops`	PEMDAS-only.
`transcendental-ops`	Contains sin, cos, pow, exp, or related operations.
`shape-ops`	Reshape, concat, transpose, squeeze, or similar operations.
`linalg-ops`	Contains `xp.linalg` or solver-like operations; `dot` is okay.
`fancy-ops`	min, max, and, or, shift, or similar operations.
`index-ops`	Contains indexing.
`nonzero-fill`	Uses a fill value other than zero.
`iterative`	Loops over a matrix or repeats until convergence.
`dense`	Exclusively dense problems.
`hypersparse`	Contains hypersparsity, such as `nnz << n` for a dimension.
`dynamic-sparsity`	Sparse-sparse interactions may change the sparsity pattern.