Formulation Machine Learning GPU Benchmark

Choosing the right hardware can reduce the time it takes to train machine learning models in the Formulation Machine Learning Panel. This page provides GPU performance benchmarks to help you select hardware for your machine learning workflows.

Any GPU provides a speedup over training on CPU alone. The size of the speedup depends on the machine learning method used, and not all methods can run on a GPU. The benchmark below trained Morgan Fingerprints with Dense Neural Network models and graph-based representation models, which are top-performing models available in the panel.

The benchmark system was a training set of temperature-dependent viscosity data for single or binary solvents, containing approximately 27,000 data points.

Benchmark Results

The benchmark measures the time required to train a model on the full training set, using 1, 4, or 12 CPU cores, with and without a GPU.

Figure 1: Training time in hours for a Formulation Machine Learning model trained on approximately 27,000 binary solvent viscosity data points, using CPU only or various NVIDIA GPUs. Lower values indicate better performance. Benchmarks were run with release version 25-4.

Based on these results, the L4 and RTX 4000 SFF Ada GPUs currently provide the best performance relative to cost. The H100 is among the fastest GPUs in this benchmark, but is not recommended for this workflow because of its high cost and its 80 GB of GPU memory, which is more than this workflow requires.