An agentic workflow for translating CUDA programs into CSL passed 26 of the 28 core tasks in a wafer-scale-system benchmark, according to an arXiv preprint. A single generation attempt passed six. Among the 26 programs that passed with the full workflow, 22 matched or beat their same-target references and 16 were strictly faster.
The task is not presented as a simple language conversion. The paper frames CUDA-to-CSL porting as architectural remapping rather than syntax substitution. Its FABRICA-Bench contains 49 paired CUDA-to-CSL tasks, while the fixed core comparison uses 28 tasks.
The workflow added checks and repair
The test compared a stripped-down attempt with a fixed-order workflow. The single-attempt arm disabled retrieval, review, repair and optimization, while the full arm generated three candidates before bounded repair.
Each candidate was copied into a fresh clone of the reference bundle and tested by the bundle's unmodified scripts. That firewall prevented the agent from changing how it was evaluated. Because the full comparison changed several workflow components together, the result does not isolate the contribution of any one component.
Different tests told different parts of the story
With the executable workflow fixed, the model comparison produced a wide spread. Opus 4.8 passed 26 of 28 core tasks, Sonnet 4.6 passed 25, Opus 4.6 passed 24, and gpt-oss-120b passed two. The last was the best listed open-weight model. The authors describe this as a preliminary capability-gap comparison rather than a statistical ranking of frontier models.
The narrower ablation examined target-specific retrieval. On a 15-task Level 1 to 3 panel, with Sonnet 4.6, a reviewer and a 20-attempt budget held fixed, the full-retrieval condition had seven passes, compared with one when the retrieved material was removed. The paper reports no confidence intervals or inferential tests for the comparison.
Across all 49 swept tasks, 38 produced at least one correct program, a task-level coverage figure of 78%. Because a task was counted once if any scheduled workflow passed, that percentage is not an independent-run pass rate.
A three-seed follow-up of the final three tasks passed eight of nine runs, and every passing run beat its reference. GEMM-1PE and Laplacian2D-Reduce passed all three seeds; LorenzoPredictor-Tile passed two of three. Only the final three tasks received this three-seed follow-up.
Speed followed correctness
Performance was checked against same-target references. Among 27 generated/reference pairs with comparable device-internal timing, the geometric-mean speedup, a multiplicative average, was 3.75 times on the simulator and 3.47 times on WSE-3 hardware. On hardware, 23 of 27 generated programs were faster, while 26 matched or beat their references.
Optimization was tested separately, starting from correct CSL. All 14 measured references improved, with a geometric-mean speedup of 1.44 times; seven programs improved by at least 25%. That result measures the tuning stage after correctness, not the full translation task.
The feedback results were more mixed. In a design-record comparison, unguided success was 21 of 40 runs without records and 20 of 40 with them. Guided translation-stage success was 33 of 40 without records and 32 of 40 with them, while measured correctness remained effectively unchanged. On Laplacian2D, however, execution-guided repair passed five of five runs, compared with zero of five in the unguided condition.
The evidence still has a narrow reach
The main caution is scope. The study uses 49 paired tasks, with 28 in the fixed core comparison, and the supplied analysis says the findings may not transfer to other kernels, architectures, SDK versions or models. The hardware comparison is narrower than the benchmark and covers 27 timing-comparable pairs. The supplied analysis reports no confidence intervals or inferential hypothesis tests.
An offline supervised fine-tuning pilot improved all four proxy metrics for a 14B model, but the manuscript says executable CSL validation is still required. Funding is not reported in the supplied text; the acknowledgment says OpenAI Codex and Anthropic Claude assisted language editing and that the authors verified the technical material. The manuscript is an arXiv preprint, version 1, dated 25 August 2026.
Paper data and sources
Original title: FABRICA: Agentic CUDA-to-CSL Translation and Optimization for Wafer-Scale Systems
Authors: Yuebo Luo, Eliu Huerta, Venkatram Vishwanath et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text