Seamlessly scale large language models with an improved software platform
Sunnyvale, Calif.– (BUSINESS WIRE)-Celebrus Systems, a pioneer in high-performance artificial intelligence (AI) computing, today extends support for PyTorch and TensorFlow on the Celebrous Software Platform (CSoft) version 1.2. Additionally, Celebras’ weight diffusion technology made it possible to quickly and easily train models with billions of parameters.
PyTorch is the epitome of machine learning frameworks. PyTorch is used by developers to accelerate deployment from research prototyping to production. As model sizes increase and transformer models become more prevalent, machine learning practitioners will need to take advantage of fast and easy-to-configure computational solutions such as the Cerebras CS-2. By running CSoft on CS-2, the developer community has access to powerful tools that enable new breakthroughs in AI.
“From the beginning, our goal was to seamlessly support the machine learning frameworks our customers wanted to write,” said Emmad Barsom, senior director of AI frameworks at Celebras Systems. Written in TensorFlow and Pytorch, our software stack, CSoft, allows you to quickly and easily represent your model in any framework you choose. You’ll have access to 2,850,000 AI-optimized cores and 40 gigabytes of on-chip memory. ”
Celebras CS-2 is the world’s fastest AI system. It is equipped with the largest processor ever, the Cerebras Wafer-Scale Engine 2 (WSE-2). The Cerebras WSE-2 has more AI-optimized compute cores, faster memory, and fabric bandwidth than any other deep learning processor in existence. CSoft is purpose-built for AI work, allowing machine learning experts to build models in open source frameworks like TensorFlow and PyTorch and run them directly on Cerebras CS-2. In fact, models written for GPUs and CPUs can be run on the Cerebras CS-2 without modification under CSoft. With CS-2 and CSoft, you can easily scale from small models like BERT to larger existing models like GPT-3.
Larger models demonstrate industry-leading accuracy in many language processing and comprehension tasks. Training these large models using the GPU can be difficult and time consuming. When learning from scratch with a new set of data, large groups of legacy equipment often require weeks and tens of megawatts of power. Additionally, as the cluster size increases, the power, cost, and complexity increase exponentially. Programming a bunch of graphics processing units requires rare skills, various machine learning frameworks, and specialized tools, each of which takes weeks of engineering time to iterate.
CS-2 was created to address these issues directly. Even the largest models can be configured in minutes, and the CS-2 performs faster than a cluster of hundreds of graphics processing units. CS-2 lets you explore more ideas in less time by spending less time on installation, configuration and training.
Celebras has customers in North America, Asia, Europe and the Middle East, GlaxoSmithKline, AstraZeneca, TotalEnergies, nference, Argonne National Laboratory, Lawrence Livermore National Laboratory, Pittsburgh Supercomputing Center, Edinburgh Parel Compute. growing number of enterprise, government and high-performance computing customers, such as the Ring Center and Tokyo Electron Devices.
For more information on the Cerebras software platform, please visit https://cerebras.net/software/.
About Celebras Systems
Celebrus Systems is a team of pioneering IT architects, computer scientists, deep learning researchers, and engineers of all types. We have come together to build a new class of computing systems with the sole purpose of accelerating AI and permanently changing the future of AI work. Our flagship CS-2 system is powered by the world’s largest processor, the 850,000-core Cerebras WSE-2, which makes your deep learning tasks much faster than GPUs.
The official version of this press release is the original language version. The translated version is provided for the convenience of the reader and has no legal effect. When using the translated version as material, please compare it with the original version, which is the only version with legal effect.