During its Ignite conference, Microsoft unveiled two custom-designed silicon chips aimed at accelerating AI and enhancing cloud workloads. The first chip is the Azure Cobalt 100 CPU, featuring a 128-core design and a 64-bit Armv9 instruction set. This cloud-native chip is set to be integrated into Microsoft's offerings and is expected to deliver up to a 40% performance boost compared to current Arm servers on Azure cloud. The Azure Cobalt 100 CPU utilizes Arm's Neoverse CSS platform, customized specifically for Microsoft.
The second chip, known as Maia 100 AI accelerator, is a key player in the AI acceleration space for running large language models. Manufactured on TSMC's 5 nm process, it boasts 105 billion transistors and supports various MX data formats, including those smaller than 8-bit, to maximize performance. While performance figures and comparisons with competing hardware from NVIDIA and AMD are yet to be released, the Maia 100 is currently being tested on GPT 3.5 Turbo. With an aggregate bandwidth of 4.8 Terabits per accelerator, the Maia 100 utilizes a custom Ethernet-based networking protocol for scaling. These chips are expected to be deployed in Microsoft data centers in early 2023, and performance benchmarks are eagerly awaited.
