2022年12月16日

作为Cerebras Cloud的唯一欧洲供应商，Green AI Cloud在一个易于使用的AI和ML应用解决方案中提供AI超级计算，数据科学和模拟工作负载Cerebras Systems是高性能人工智能(AI)计算的先驱，而Green AI Cloud是欧洲最可持续的超级计算平台，宣布在Green AI上提供Cerebras Cloud。作为欧洲第一家提供业界领先的CS-2系统的云计算提供商，Green AI客户现在可以轻松地训练gpt级模型，速度比传统云服务提供商快得多，对环境的影响也大大减少。“Cerebras致力于通过云、内部或混合方式，以最环保的方式为全球客户提供人工智能计算。”绿色人工智能云是清洁能源计算的先驱。绿色AI云在其位于瑞典北部的数据中心仅使用可再生能源(水电和风能)，并将操作中的多余热量转化为工业产品制造的加热液体，从而实现负的二氧化碳足迹。凭借其新的Cerebras Cloud产品，客户现在可以训练领先的生成式变压器(GPT)级模型，包括GPT- j、GPT-3和GPT- neox，并承诺通过Green AI独特的可持续和“碳智能”云实现等效的碳抵消。Green AI Cloud创始人兼首席执行官Jacob Bostrom表示:“我们很高兴为客户提供行业领先的CS-2系统，该系统由85万颗AI核心WSE处理器提供支持，这将大大减少他们最初的编程工作，并轻松与现有工作流程集成。”“通过与Cerebras的合作，我们提供了最新、最快的人工智能技术，我们不仅继续推动人工智能创新的未来，还将继续推动可持续发展和环境责任的下一个时代。”对于包括GPT-J、GPT-3和GPT-3XL在内的数十亿参数NLP模型，Cerebras CS-2的能源效率是NVIDIA®A100的3.5倍。对于人工智能模型“BERT Base”，Cerebras CS-2的能源效率是NVIDIA V100的6倍(根据马萨诸塞大学阿默斯特分校和麻省理工学院技术评论)。使用绿色AI云与其他欧洲云提供商相比，在训练单个深度学习AI变压器模型时，所产生的碳抵消相当于11000棵树每年吸收的二氧化碳! Safeguarding Data Security and Privacy Regulations The EU has some of the strictest data privacy and security rules in the world. A cloud provider based in the EU, such as Green AI Cloud, enables customers across the EU to benefit from Cerebras’ industry-leading AI compute and stay within the data privacy structures. Customers across segments, from pharmaceutical to finance services, and energy to heavy manufacturing, now have push-button access to the Cerebras CS-2 system and can use it by the day, week, or month. “Cerebras is committed to delivering AI compute – via cloud, on premise or hybrid – to our customers around the world in the most environmentally efficient way possible,” said Andrew Feldman, co-founder and CEO, Cerebras Systems. “As the leader in energy efficient AI compute, it was an obvious choice to partner with and deliver AI compute to the Green AI Cloud.” Cerebras recently announced its AI Model Studio, as well as partnerships and wins with Sandia National Laboratory and generative AI pioneer Jasper. At the NeurIPS Conference earlier this month, Cerebras broke new ground demonstrating high sparsity training on GPT class models, showing that GPT models can be trained at 90% sparse, using a fraction of the FLOPs and time, and still achieve the same accuracy.