An AI agent helped move 162 computational sections of the CReSS weather model to a graphics processing unit; verification found five numerical discrepancies
An AI agent helped move 162 computational sections of the CReSS weather model to a graphics processing unit; verification found five numerical discrepancies
On August 13, the authors of a preprint described how they used an AI agent to move 162 of 387 code sections in CReSS, a cloud and storm model, to a graphics processing unit. In a 30-minute typhoon simulation, the version that ran these sections on the graphics processing unit was 5.1 times faster than the original version running on the central processing unit.
CReSS has been under development since 1998 for simulations of tropical cyclones, heavy rainfall, and convective systems. In this type of model, different parts of the program pass their results to one another. Moving them to a graphics processing unit can change the order of parallel operations, rounding, and sometimes the output of a built-in mathematical function. The authors therefore measured the performance gain and checked whether the model preserved its behavior.
For each of the 162 sections, the team recorded the actual input data and the output of the standard central processing unit version during a 360-step simulation. Using this snapshot, the AI agent prepared a separate test program and a version containing OpenACC directives, which tell the compiler how to parallelize the computation on the graphics processing unit. The output from the old and new versions was compared with the recorded value for every number. The team examined any relative difference greater than one hundred-thousandth separately.
Five sections each produced one value above this threshold. In one section, the central processing unit calculated a temperature of 233.16002 K, while the graphics processing unit calculated 233.16000 K. The difference in the final digit placed the values on opposite sides of the 233.16 K threshold. As a result, the graphics processing unit version applied a latent heat correction, which accounts for the energy that water releases or absorbs when it changes phase. The CReSS developers considered these differences acceptable for this scenario.
“The test using the recorded snapshot did not reveal the error. It appeared only after integration,” the authors write.
The full simulation reached another call to this section, where a program branch for a specific condition was executed. The authors reproduced this branch in the test and compared the two versions again. In the full simulation, the maximum and minimum pressure perturbations differed by 1.0×10⁻⁵ and 5.6×10⁻⁵, respectively. Both values were below the 10⁻⁴ criterion.
The AI agent prepared the code transformations and tests, while the developers decided whether each numerical difference was acceptable.