Fournex

Fournex is a GPU profiling tool designed for engineers seeking evidence-based performance analysis and optimization. It addresses the challenge of identifying and resolving complex GPU performance issues by providing guided profiling, concrete diagnoses, and actionable recommendations. The tool operates through a command-line interface, starting with the 'frx init' command to check the user's machine, detect the GPU model, and suggest the appropriate profiling command for CUDA kernels or PyTorch training. It supports additional frameworks such as JAX and TensorFlow, and works with NVIDIA GPUs, including models like H100, A100, L4, RTX4090, and RTX5060.

Fournex collects lightweight runtime telemetry, Nsight Compute (NCU) counters, PTX structure, and source code information to identify and rank GPU bottlenecks. Its analysis detects specific issues such as uncoalesced global loads, L1 and L2 cache thrashing, low occupancy due to register pressure, scheduler underutilization, and host or input stalls. The tool distinguishes between different types of memory pressure and provides targeted recommendations, including actions to coalesce memory loads, reduce working sets, tune occupancy, or address framework overhead. Each finding is supported by the exact metric, threshold crossed, confidence level, and validation steps, allowing engineers to verify improvements and understand the risks before making code changes.

The profiling process is guided and interactive, with the ability to generate opportunity-ranked lists of kernels, highlight framework abstraction overhead, and suggest next steps with validation commands. Fournex can produce optimization briefs suitable for use with large language models by adding the '--explain' flag during profiling. Reports are readable in the terminal and maintain a traceable evidence trail from diagnosis to benchmark proof, which can be used for code review or further validation.

Fournex is aimed at engineering teams working on compute-intensive workloads and is trusted by organizations in the AI and machine learning space. The tool focuses on explainability, repeatability, and evidence-backed recommendations, emphasizing measurable improvements and guarded automation rather than blind optimization.