PallasBench: Robust JAX Pallas GPU Kernel Benchmarks for A100

PallasBench is a suite of 45 JAX Pallas kernels across 3 difficulty levels, originally designed for TPU and adapted for GPU. The dataset reports that 39 out of 45 kernels pass on an NVIDIA A100 80GB, representing the first GPU-focused evaluation of JAX Pallas kernels. It was authored by EvanOLeary and last updated on May 30, 2026.

Use Cases

• Benchmarking GPU compatibility of JAX Pallas kernels based on the suite of 45 adapted kernels.
• Comparing kernel performance across difficulty levels based on the described three-tier structure.
• Validating fixes for Triton compiler block size limits on NVIDIA hardware, as mentioned in the description.

Strengths

• Contains 45 distinct JAX Pallas kernels for evaluation.
• Reports a specific pass rate of 39 out of 45 kernels on a defined hardware platform (NVIDIA A100 80GB).
• Kernels are categorized across 3 difficulty levels, allowing for tiered analysis.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

huggingface

Collection Method

Likely contains benchmark results from executing and adapting JAX Pallas kernels.

Freshness

Last updated 2026-05-30 06:07:30; freshness should be verified.