AWS is already Antropic’s main cloud service provider. In the future, Antropic plans to use Tranium and Inferentia chips to train and deploy basic models. Antropic also plans to cooperate in the development of Tranium through a ‘hardware-software development approach.’
It’s unclear whether the agreement requires Antropic to use AWS chips exclusively, but it’s clear it’s a move by Amazon to challenge Nvidia and other dominant players as the AI chip race accelerates.
Forrester senior analyst Alvin Nguyen said, “This collaboration is the first step toward expanding the accessibility of generative AI and AI models.”
Accelerating Claude Development
Antropic, founded in 2021, has made great progress this year by competing with OpenAI through Claude’s LLM. The Claude 3 model family consists of Sonnet, Haiku (the fastest and most compact model), and Opus (a model suitable for complex tasks), all of which are available on Amazon Bedrock. These models feature vision capabilities and a 200,000 token context window, allowing them to process approximately 150,000 words or 500 pages of data.
In particular, last October, Antropic added the “Computer Use” function to the Claude 3.5 Sonnet model. This feature allows the model to use the computer like a human, quickly performing tasks such as moving the mouse cursor, switching tabs, navigating websites, clicking buttons, and compiling research documents. Antropic claims that the Sonnet model outperforms all other currently available models for agentic coding tasks.
According to AWS, adoption has spread rapidly since Claude was added to Amazon Bedrock, a fully managed service for building generative AI models, in April 2023. It is currently used by “tens of thousands” of companies across a variety of industries. This foundation model is used to build a variety of functions such as chatbots, coding assistants, and complex business processes.
“2024 has been a year of dramatic growth for Claude, and our collaboration with Amazon has been instrumental in bringing Claude’s capabilities to millions of end users through Amazon Bedrock,” said Dario Amodei, co-founder and CEO of Antropic, in a statement. “He said.
This expansion of cooperation has strategic significance for both companies. It demonstrates that Antropic’s model has high performance and versatility, and suggests that AWS’s infrastructure can handle generative AI workloads powerful enough to compete with other chip manufacturers such as NVIDIA.
According to Forrester’s Nguyen, the collaboration provides Antropic with “the assurance of infrastructure and the ability to continuously expand and roll out model capabilities” and also helps expand Antropic’s presence and accessibility.
“This collaboration shows that Antropic can work effectively with a variety of partners,” Nguyen said. This increases confidence in our ability to train, create models, and utilize them. AWS evaluated that through cooperation with Antropic, “We have secured a representative customer in the AI field.”
Optimizing all aspects of model training
As part of the expanded collaboration, Antropic will participate in the development and optimization of future versions of AWS’ custom ML chip, Trainium. Trainium supports deep learning learning for models with more than 100 billion parameters.
Antropic said it is working closely with AWS’s Annapurna Labs to write a low-level kernel that can interact with Tranium silicon. We are also contributing to the AWS Neuron software stack to power Tranium and are collaborating with the chip design team to improve hardware computational efficiency.
“By combining our close hardware-software co-development approach with the strong price-performance and massive scalability of the Tranium platform, we can optimize all aspects of model training from silicon to the entire stack,” Antropic said in a blog post. He said.
“This approach offers an advantage over commodity hardware like NVIDIA GPUs that do more than is absolutely necessary,” Forrester’s Nguyen said. “In addition, the long-standing cooperative relationship between the two companies has the potential to offset to some extent the performance optimization advantage provided by the NVIDIA CUDA platform,” he emphasized.
“Deep collaboration between software and hardware engineers and developers allows us to achieve levels of optimization across both hardware and software that would be difficult when working independently,” Nguyen added.
editor@itworld.co.kr
Source: www.itworld.co.kr
