Rebecca Lewington, Technology Evangelist | March 21, 2022

We just published a case study which explores a project we conducted with a leading financial services institution to help them overcome a roadblock to using advanced neural network models for a wide range of natural language processing (NLP) tasks.

Actually, we were able to do more than “help”. Our CS-2 system delivered the compute performance of more than 120 AI-optimized GPUs. No marginal gains here; that’s a huge leap.

With that performance, we were able to reduce training time for a complex BERTLARGE model by 15X, compared to a leading 8-GPU server, demonstrate dramatic improvements in model prediction confidence, while almost halving energy consumption.

Those are compelling results in any industry, but especially so in the ultra-competitive world of finance, where even small gains in performance can lead to huge monetary rewards. Which explains why the customer asked to remain anonymous. In this field, it isn’t just time that is proverbially money, but also technology.

The case study, “Accelerating NLP Model Training and Enabling Higher Accuracy for Financial Services Applications” was authored by a trio of Cerebras’ women: machine learning solutions engineers Sanjana Mallya and Cindy Orozco Bohorquez, along with machine learning product director Natalia Vassilieva.

The exciting thing for the customer is that this is about more than simply training an existing model faster. It’s about giving them the freedom to rapidly experiment to find better models, which they couldn’t explore previously because training those new models took too long to be useful. Data scientists at financial institutions have better things to do with their time than waiting for their models to train themselves.

What do I mean by “better”? Large neural language models, such as BERT, are really good at many natural language processing tasks. But models trained on generic data, such as text extracted from web pages, tend not to work very well when asked to work with “domain-specific” text, laden with, say, technical terms from the world of finance. They don’t get the lingo, so to speak.

A common approach to re-educate a generic model is to “fine-tune” it, which means to run additional training using a domain-specific dataset. However, past work has shown that for specialized domains with abundant domain-specific texts, you can get more accurate models by skipping the generic training and training from scratch (Jinhyuk Lee, et al 2020). But there’s a problem: fine-tuning is relatively easy with conventional computing systems. Training from scratch, on the other hand, is very computationally intensive. It takes a long time, evening using large clusters of servers stuffed with GPUs, looked after by specialist teams. As are result, training from scratch has remained out of reach for most companies.

Cerebras can fix that, making training from scratch quick and easy for mere mortals to do. The results of this report show a promising path to accelerate research and AI-powered capability development for financial services enterprises, but those results are equally applicable anywhere NLP models are used, including healthcare, life sciences, energy, manufacturing, cybersecurity and government services.


For more information about the Cerebras CS-2 system and its financial services applications, please visit

Read the case study here.

If you would like to evaluate how the CS-2 system can benefit your organization, we encourage you to get in touch here.