SASS-Bench: Hardware-Grounded GPU Execution Prediction Benchmark

50 compiled SASS kernels with 2-3 input-regime variants each form a benchmark for predicting exact register values and output buffers. Ground truth was captured from a Blackwell sm_120 GPU using NVBit dynamic binary instrumentation. The dataset was created by anon-neurips-ed-3552 and last updated on 2026-05-07.

Use Cases

• Training models to predict exact uint32 register values at specified instructions based on compiled SASS kernels.
• Benchmarking model accuracy for predicting the final output buffer bit-patterns given raw inputs and launch configuration.
• Developing hardware-grounded AI systems for GPU assembly code analysis and performance estimation.

Strengths

• 50 distinct compiled SASS kernels provide a foundation for model evaluation.
• Ground truth captured via NVBit dynamic binary instrumentation on a specific Blackwell sm_120 GPU.
• Includes 2-3 input-regime variants per kernel, suggesting controlled experimental variation.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• The total number of rows and data file formats are unknown, which may limit suitability assessment.

Provenance

Source

huggingface

Collection Method

Ground truth captured via NVBit dynamic binary instrumentation on a Blackwell sm_120 GPU.

Time Range

null

Freshness

Last updated 2026-05-07 08:59:49; freshness should be verified.

Geography

null