Cerebras

Cerebras is a platform for fast and easy AI training. It includes AI supercomputers, model services and cloud options to speed up large language model and other AI work.

Cerebras is centered on its wafer-scale engine (WSE-3) processor, which has 900,000 AI processing cores, 44GB of onboard memory and 21PB/s of memory bandwidth. It's designed to deliver the performance of a cluster of machines in a single chassis, a design that's well suited to high-performance computing tasks like training large language models.

The platform includes a range of tools and services to help with AI work:

• AI Model Services: Work with Cerebras experts to develop and train generative AI models tailored to your business needs, such as custom chatbots, code completion and summarization.
• Cloud Solution: Run AI supercomputer performance in the cloud so you can train large models without the hassle of setting up your own infrastructure.
• Software Platform: The Cerebras Software Platform, CSoft, is designed to work with frameworks like PyTorch so researchers can easily move their models to the Cerebras system. It includes tools for writing custom kernels and optimizing performance.
• Cerebras AI Model Studio: A pay-by-the-model computing service for training and fine-tuning large language models, with deterministic performance and push-button simplicity.

Cerebras is suited for industries like Health & Pharma, Energy, Government, Scientific Computing and Financial Services, where large-scale AI work is more common. Customers include National Laboratories, Aleph Alpha, The Mayo Clinic and GlaxoSmithKline, which have seen a significant reduction in training times and improvement in model accuracy using Cerebras.

Pricing is flexible, with customers able to pay by the hour or by the model, depending on the needs of the project. That flexibility means customers can pay for AI development work according to their budget and goals.

By offering an optimized platform for AI development, Cerebras lets researchers and businesses explore new possibilities with large language models and high-performance computing.

Published on July 12, 2024