High performance AI chip maker Cerebras, maker of the Wafer-Scale Engine, the world’s largest computer chip, today announced Cerebras Cloud @ Cirrascale, delivering its AI accelerator as a cloud service. The alliance combines Cerebras’s CS-2 HPC/AI system, featuring 850,000 AI-optimized compute cores, with Cirrascale’s deep learning cloud services infrastructure.
“With Cerebras Cloud @ Cirrascale, (the CS-2) is available right at your fingertips,” said Gil Haberman, Cerebras senior director of product marketing. “Cerebras Cloud is available in weekly or monthly flat-rate allotments, and as you grow there are discounts offered for longer-term, predictable usage. In our experience, as users observe the blazing fast performance of the CS-2, ideas for new model and experiments emerge – such as training from scratch on domain-specific datasets, using more efficient sparse models, or experimenting with smaller batch sizes – resulting in better-performing models in production and accelerated pace of innovation.”
Cerebras was one of the early entrants in the growing AI chips field – processors built specifically for machine learning model training rather than general purpose chips placed in model training workloads. Running contrary to the tech industry’s customary drive for miniaturization, Cerebras’s unusually large, dinner plate-sized chips have drawn wide attention, as seen in a portrait of Cerebras published in The New Yorker last month.
“Nearly every day, we engage with machine learning (ML) scientists and engineers who are looking to push the frontiers of deep learning but find themselves constrained by long training times of existing offerings,” Haberman said. “In contrast, our solution has been built from the ground-up for AI. It delivers hundreds or thousands of times more performance than alternatives – enabling data scientists and ML practitioners to train and iterate on large, state of the art models in minutes or hours rather than days or weeks.”
Haberman said Cerebras Cloud offering with Cirrascale expands the company’s reach, “an important step in truly democratizing high-performance AI compute.”
Cerebras Wafer-Scale Engine
“The CS-2 systems that power Cerebras Cloud deliver cluster-scale performance with the programming simplicity of a single node,” he said. “Whether the model is large or small, our compiler optimizes execution to get the most out of the system. As a result, cluster orchestration, synchronization and model tuning are eliminated, letting you focus on innovation rather than cluster management overhead.”
The software platform integrates with popular machine learning frameworks, such as TensorFlow and PyTorch, “so you can use familiar tools to get started running models on the CS-2 right away. The Cerebras Graph Compiler automatically translates your neural network from your framework representation into a CS-2 executable, optimizing compute, memory, and communication to maximize utilization and performance,” Haberman said.
He added that for customers with data stored in other cloud services, Cirrascale can integrate the Cerebras Cloud with other cloud-based workflows to create a multi-cloud solution.
