- Microsoft is set to introduce an Arm-based device for general computing tasks alongside its earliest artificial intelligence chip.
- During its Burn meeting in Seattle, Microsoft revealed that both devices will leverage its Azure cloud platform.
- Amazon Web Services’ Graviton Arm microprocessor, initially unveiled five years ago, has gained widespread adoption.
At a Microsoft facility in Redmond, Washington, the Maia 100 AI Accelerator and its companion are housed on a custom-built plate. The companion manages the computational requirements of AI workloads by circulating liquid to and from the rack, akin to a vehicle’s radiator.
Microsoft unveiled two chips at its Ignite conference in Seattle on Wednesday.
The first chip, the Maia 100 artificial intelligence chip, is positioned to rival Nvidia’s highly coveted AI graphics processing units. The second chip, named Cobalt 100 Arm chip, is designed for general computing tasks and poses competition to Intel processors.
Major technology firms are expanding the range of cloud infrastructure options available to customers for running applications. Companies like Alibaba, Amazon, and Google have been offering these services for years. Microsoft, boasting approximately $144 billion in cash reserves as of October, held a 21.5% share of the cloud market in 2022, trailing only Amazon according to estimates.
In an interview with CNBC, Rani Borkar, a commercial vice president, mentioned that virtual-machine scenarios utilizing Cobalt cards will be commercially accessible through Microsoft’s Azure platform in 2024. However, she did not provide a specific release date for the Maia 100 chip.
Google introduced its original tensor processing unit for AI back in 2016. Amazon Web Services unveiled its Graviton Arm-based chip and Inferentia AI processor in 2018, followed by the Trainium chip for model training in 2020.
Specialized AI chips from cloud providers could help alleviate the GPU shortage. Unlike Nvidia or AMD, Microsoft and other cloud computing giants do not intend to sell servers containing their chips.
Borkar highlighted that Microsoft developed its AI technology based on feedback from customers.
Microsoft is currently evaluating how the Maia 100 chip aligns with the requirements of its Bing search engine’s AI bot, the GitHub Copilot programming assistant, and GPT-3.5 Turbo, a substantial language model supported by Microsoft-backed OpenAI. These language models, trained on vast internet data, can generate email responses, summarize documents, and provide concise answers to queries.
Colette Kress, Nvidia’s finance chief, mentioned at an Evercore conference in New York in September that collaborative efforts were underway to enhance supply chain resilience and meet customer demand.
OpenAI has recently been training models on Nvidia GPUs within Azure.
Apart from the Maia device, Microsoft also introduced Sidekicks, a series of liquid-cooled technology designed to complement Maia servers within racks. These Sidekick units can be seamlessly integrated alongside the server racks without requiring retrofitting.
Efficiently utilizing data center space with GPUs can be challenging. According to Steve Tuck, CEO of Oxide Computer, some businesses strategically place GPU servers at the bottom of racks to prevent overheating, rather than filling the cabinet from top to bottom. Additional cooling mechanisms are sometimes incorporated to maintain optimal temperatures.
Transitioning from GPUs to AWS Trainium AI cards may pose challenges compared to switching to Intel Xeons or Gravitons due to unique imperfections in each AI design. However, over time, comparable price-performance benefits are anticipated with Trainium chips compared to GPUs.
Borkar emphasized the importance of sharing specifications with the ecosystem to benefit all cloud customers.
Regarding the effectiveness of Maia compared to competing models like Nvidia’s H100, Borkar mentioned a lack of detailed information. Nvidia’s announcement of the H200 shipping in the second quarter of 2024 indicates ongoing advancements in the field.
