Choosing Cloud or Local GPUs for Deep Learning Trading Research
Summary
The article explains why GPU acceleration can shorten deep learning training and compares renting cloud compute with buying a local workstation for quantitative research. It frames the decision around model and parameter-search demands, electricity costs, existing data infrastructure, personal workflow, and restrictions on cloud use. Cloud instances can be started for intensive training and shut down afterward, while trained models may run on less costly hardware. Local machines require more upfront spending but offer customizable storage and avoid ongoing rental charges beyond electricity.
The discussion uses historical AWS P2 pricing and example Nvidia GPU prices to illustrate the cost comparison, while noting that prices and performance change quickly and that storage and data transfer add to cloud costs. It also describes practical tradeoffs: rented instances are headless Linux servers accessed through SSH, whereas local workstations require assembly and software setup, including CUDA and machine learning libraries. The article recommends starting with rented resources when requirements are uncertain, then tailoring a purchase to a defined research plan. Its figures are time-specific, and it does not provide a general cost model or benchmark across vendors.
Key ideas
- GPU acceleration can reduce deep learning training times and support faster quantitative research iteration.
- Cloud rental shifts costs to usage and can be stopped after model training, though storage and data transfer may add expense.
- A local workstation requires upfront investment but allows hardware customization and local data storage.
- The better choice depends on workload, search scope, electricity costs, data setup, workflow, and cloud restrictions.
- The article’s vendor prices and hardware examples are historical and may no longer reflect current costs.
Tags
This summary was written by Stratmill's research agent from the original; it is not a copy of the source.