On July 26th, according to Reuters' latest report, six Amazon AWS engineers conducted tests on a newly protected server at a chip laboratory in Austin, Texas.
It is worth mentioning that this server is equipped with an artificial intelligence chip that Amazon AWS is independently developing. Although Amazon did not disclose specific information such as the name and parameters of the chip used in the test server, according to David Brown, Vice President of Computing and Networking at AWS, "In some cases, the performance of this chip can be improved by 40% to 50% compared to NVIDIA, so its cost should be half that of running the same model with NVIDIA
From the perspective of Amazon's business matrix, the AI specific chips previously developed by the company's AWS cloud computing service platform include Inferentia and Trainium. According to Amazon, during the recent Prime Day (Amazon Prime Day, the platform's promotional day), the company deployed 80000 AI chips to implement cloud computing in response to the surge in activity on its platform.
It is not difficult to understand that such a large investment is driven by the profit of Amazon AWS business. Data shows that AWS's sales account for nearly one-fifth of Amazon's overall revenue, and in the first quarter of this year, its sales increased by 17% year-on-year, reaching $25 billion. Currently, AWS holds nearly one-third of the cloud computing market share.
Therefore, it is evident that as the main growth driver, Amazon has a continuous demand for cost reduction and increased computing power in its cloud computing business. The purpose of testing the server this time is also very simple. During a visit to the laboratory, company executive Rami Sinno bluntly stated that it is to "compete with market leader NVIDIA's chips
NVIDIA achieves the 'target of public criticism'
Xinuo pointed out that Amazon's customers are increasingly seeking a lower cost alternative that can handle complex calculations like Nvidia AI chips in terms of performance. In fact, since Nvidia won the "throne" in the AI chip field, many manufacturers have taken action to get rid of their dependence on it and the high "Nvidia tax".
Leading the charge, such as Microsoft, was reported in February this year to be developing a new network card to enhance the performance of Maia AI server chips. At that time, the banner of the project was "reducing the time spent by OpenAI training models on Microsoft servers while lowering training costs
Soon, more and more manufacturers chose to join the self-sufficient "de Nvidiaization" camp. In April of this year, Google launched the Arm architecture AI chip Axion, which can support search engine operation and AI related tasks. In addition, Meta has also released the latest version of its self-developed AI chip MTIA, stating that its goal is to reduce reliance on companies such as Nvidia.
Market research firm CFRA analysts believe that large technology companies are facing pressure on chip costs and need to rely on self-developed chips to alleviate it. In fact, for large technology companies with strong financial strength, self-developed AI chips can not only break free from the monopoly effect of leading manufacturers such as Nvidia, but also customize hardware that better meets their own needs according to their own business.
While other manufacturers are vigorously promoting the "go Nvidia" movement, there are also manufacturers like Broadcom brewing to become the "second Nvidia". According to Broadcom's first quarter financial report, the company expects AI revenue to exceed $10 billion in the first quarter of 2024, with customized DPU chips accounting for approximately $7 billion. Just a few days ago, it was reported that the company was in talks with OpenAI to develop customized new AI chips for the latter.