Datasetpublic

harry1332/hpca2027-agentic-validated-traces-20260711

HPCA 2027 CPU-Local Agentic Validated Traces This dataset contains validated CPU-local traces used for HPCA 2027 agentic workload characterization. AppWorld, CORE-Bench, and Terminal-Bench are paper-backed benchmark workloads from official repositories; the hybrid-RAG and data-analysis agents are deterministic characterization workloads retained from the initial pass. Workloads And Full 100M Results Workload Source Frontend Backend Bad speculation Retiring… See the full description on the dataset page: https://huggingface.co/datasets/harry1332/hpca2027-agentic-validated-traces-20260711.

sourceHugging Facemitupdated 2mo agoView on Hugging Face

1likes80downloads

Dataset Card

HPCA 2027 CPU-Local Agentic Validated Traces

This dataset contains validated CPU-local traces used for HPCA 2027 agentic workload characterization. AppWorld, CORE-Bench, and Terminal-Bench are paper-backed benchmark workloads from official repositories; the hybrid-RAG and data-analysis agents are deterministic characterization workloads retained from the initial pass.

Workloads And Full 100M Results

Workload	Source	Frontend	Backend	Bad speculation	Retiring
`appworld`	ACL 2024 paper / official repo	8.19%	51.66%	19.21%	20.94%
`core_bench_tomography_3497606`	TMLR 2025 paper / official repo	4.33%	43.01%	12.55%	40.11%
`terminalbench_train_fasttext`	ICLR 2026 paper / official repo	4.35%	51.63%	11.47%	32.55%
`hybrid_rag_100k_d384_q64`	deterministic CPU hybrid retrieval workflow	0.005044%	83.782907%	0.012178%	16.199871%
`data_analysis_agent_e4m_r8`	deterministic CPU data-analysis workflow	3.298938%	54.668990%	4.322370%	37.709701%

AppWorld runs released train/development solutions against its local 457-API environment. CORE-Bench runs official Code Ocean capsule 3497606, the iterative tomography workload. Terminal-Bench runs its official train-fasttext Yelp model-training task. All three measured phases are CPU-only and exclude external model/API waiting.

Percentages are derived from the four raw Scarab top-down counters in each workload's full-simulation validation directory. All four values sum to 100%.

Bundle Layout

Every workload uses the legacy BFS/ASPLOS-compatible directory structure:

text

<workload>/
  fingerprint/
  simpoints/
  traces/whole/<dr-folder>/{raw,trace,bin}/
  traces_simp/{raw,trace,bin}/
  trace_clustering_info.json
  validation/
  workload/

AppWorld selected segments 4, 5, and 9 with weights 0.100654, 0.100654, and 0.798692. CORE-Bench selected segments 3, 7, and 8 with weights 0.300328, 0.499453, and 0.200219. Terminal-Bench selected segments 2, 5, and 6 with weights 0.376507, 0.376507, and 0.246985.

The AppWorld and CORE-Bench traces require decoder history. Their selected ZIPs therefore contain contiguous chunks 0..selected, while retaining the same four-digit chunk names and BFS directory/descriptor interface. Simulations must use the selected segment's standard 10M-instruction ROI offset.

Validation

For AppWorld, CORE-Bench, and Terminal-Bench:

—the whole trace covers the intended 100M-instruction window after a 1B skip;
—every ZIP passes integrity testing;
—selected weights normalize to 1.0;
—every selected archive completes a 1M-instruction Scarab smoke simulation;
—selected SimPoints have distinct raw top-down vectors;
—full simulations use the Golden Cove setup with --dcache_assoc 8;
—full simulations reproduce 51.66%, 43.01%, and 51.63% backend bound;
—workload paper, repository revision, command, and compatibility notes are recorded under workload/ and validation/.

Per-file SHA-256 checksums are provided in the root SHA256SUMS. The DynamoRIO ARCH_REGDEPS warning is also present in the previously validated database and agentic releases; successful Scarab completion and raw counters are the acceptance criteria.