wtf is 3080bench?

This repository contains a benchmarking study that measures how fast a specific AI language model, Qwen3.6-27B, runs across four modified RTX 3080 graphics cards with 20GB of memory each. The cards are installed in a machine located in China. The project tests two main operations: how quickly the system processes an initial prompt, and how quickly it generates new text tokens as the conversation gets longer. It pairs these tests with a self-contained dashboard to display the results. The benchmark runs a matrix of configurations. It tests two different power limits on the cards, 220 watts and 320 watts, alongside two different model compression formats. The system splits the work across the four cards in two ways: by tensor or by layer. The tests use both synthetic data generated by a built-in tool and realistic coding prompts to simulate actual usage. The study also examines speculative decoding methods, which attempt to guess upcoming words to speed up generation. A set of scripts orchestrates the entire process. One script sets the power limit on the cards, another launches the server, and a third collects telemetry data like temperature and memory usage at half-second intervals. A main orchestration script runs all the tests in sequence and can resume if interrupted, skipping configurations it has already completed. After the tests finish, a separate analysis script processes the results into summary spreadsheets and charts. The README includes detailed notes about the specific challenges of running this setup on the chosen machine. The machine lacks direct access to standard model hosting sites, so the project pulls models from a mirror site instead. Long running jobs require specific command line flags to survive disconnections, and the CUDA toolkit needed additional development headers installed manually to build the software correctly.