On Tuesday, Amazon.com’s cloud unit unveiled new data center servers that are equipped with its own AI processors. These servers are expected to pose a challenge to Nvidia, and Apple has signed on as a customer to utilize them.
Anthropic, an AI startup, will be the first to utilize the new servers, which are constructed from 64 Trainium2 processors from Amazon Web Services. These servers will be connected to form a massive supercomputer with hundreds of thousands of chips. Additionally, Apple executive Benoit Dupin disclosed that the company is employing Trainium2 chips.
Nvidia commands over 70% of the AI processor market, and its competitors in the traditional chip industry, including Advanced Micro Devices, are making every effort to catch up.
However, Meta Platforms, Microsoft, and Alphabet’s Google are among Nvidia’s most formidable competitors, and they all possess their own custom AI processors. Meta’s chip is utilized for internal operations, while Amazon and Google market their chips to paying customers in addition to utilizing them internally.
The company’s next iteration of AI chip, Trainium3, will be introduced next year, according to AWS Chief Executive Matt Garman.
“The new offerings are specifically designed to handle the rigorous workloads of cutting-edge generative AI training and inference,” stated Garman at the event in Las Vegas on Tuesday.
Nvidia’s flagship server, which is equipped with 72 of its most recent “Blackwell” processors, will be competing with the new servers, which AWS refers to as Trn2 UltraServers. Nvidia and AWS both provide proprietary technology for connecting chips. However, Gadi Hutt, AWS’s business development manager for AI chips, stated that AWS will be able to connect a larger number of chips than Nvidia.
In an interview, Hutt stated, “We believe that Trainium2 will provide customers with a greater amount of compute than what is currently available from Nvidia, allowing them to save money.” He also mentioned that certain AI models can be trained at a 40% lower cost than on Nvidia processors.
AWS executives stated that the new servers and massive supercomputer would be operational by the end of the year, but they did not specify a specific date. Nvidia’s shipments have been restricted by supply chain constraints, despite the fact that both AWS and Nvidia are racing to release their flagship offerings in response to the increasing demand.
Both Nvidia and AWS rely on Taiwan Semiconductor Manufacturing to produce their chips.
In terms of supply, Hutt informed Reuters, “We are in a relatively good position throughout the entire supply chain.” “Trainium chips are the only item that cannot be dual-sourced when we construct the systems.”
