AI software platform enables giant model training
Version 1.2 of the company’s Cerebras Software Platform, CSoft, features expanded support for PyTorch and TensorFlow. In addition, says the company, customers can now quickly and easily train models with billions of parameters via Cerebras’ weight streaming technology.
PyTorch, the leading machine learning framework, is used by developers to accelerate the path from research prototyping to production deployment. As model size increases and as transformer models become more popular, says the company, it is essential that machine learning practitioners have access to fast, easy to set up and use compute solutions like the Cerebras CS-2 AI system.
With the CS-2 running CSoft, says the company, the developer community has a powerful tool to enable new breakthroughs in AI.
“From the start, our goal was to seamlessly support whichever machine learning framework our customers wanted to write in,” says Emad Barsoum, Senior Director, AI Framework, at Cerebras Systems. “Our customers write in TensorFlow and in PyTorch, and our software stack, CSoft, makes it quick and easy to express your models in the framework of your choice. By doing so, our customers gain access to the 850,000 AI optimized cores and 40 Gigabytes of on-chip memory in the Cerebras CS-2.”
Claimed as the world’s fastest AI system, the Cerebras CS-2 is powered by the largest processor ever built – the Cerebras Wafer-Scale Engine 2 (WSE-2). The Cerebras WSE-2 delivers more AI optimized compute cores, more fast memory, and more fabric bandwidth than any other deep learning processor in existence, says the company.
Purpose built for AI work, the CS-2 runs CSoft, which enables machine learning practitioners to write their models in the opensource frameworks of TensorFlow or PyTorch and, without modification, run the model on the Cerebras CS-2. In fact, says the company, a model that was written for a graphics processing unit or a central processing unit can run under CSoft on the Cerebras CS-2 without any changes. With the CS-2 and CSoft, practitioners can seamlessly scale up from small models like BERT to the largest models in existence like GPT-3.
Large models have demonstrated state-of-the-art accuracy on many language processing and understanding tasks. Training these large models using GPU is challenging and time-consuming. Training from scratch on new datasets often takes weeks and 10s of megawatts of power on large clusters of legacy equipment. Moreover, as the size of the cluster grows, power, cost, and complexity grow exponentially. Programming clusters of graphics processing units requires rare skills, different machine learning frameworks, and specialized tools that require weeks of engineering time to each iteration.
The CS-2 was built to directly address these challenges. Setting up even the largest model takes only a few minutes, says the company, and the CS-2 is faster than clusters of 100s of graphics processing units. With less time spent in set up, configuration and training, the CS-2 enables users to explore more ideas in less time.