Microsoft introduced the new Maia 100 AIC, which will be an important step in the development of cloud technologies. As part of the Hot Chips 2024 conference, the company revealed the details of her first generation of custom AI-accelerators specially designed to process large-scale tasks of artificial intelligence in the cloud environment Azure.
Maia 100 is a unique system integrated to achieve maximum performance when minimizing costs. It includes the architecture of the platform with custom server boards and specialized racks, as well as the software optimized to increase the efficiency of services, such as Azure Openai.
The main element of the system is a chip of about 820 square millimeters, which was created using the advanced TSMC technology and is equipped with four HBM2E chips, providing bandwidth of 1.8 terabyte per second and a capacity of 64 gigabytes. Such architecture allows you to cope with high loads associated with the processing of large volumes of data.
MAIA 100 supports various data formats and provides high performance for energy consumption of up to 700 W, which makes it an effective solution for tasks that require significant computing resources. Particular attention is paid to the support of modern machine learning algorithms, which is reflected in the architecture of processors and blocks of tensor operations.
The system is also equipped with software tools that allow you to effectively manage resources and reduce energy consumption. For this, data compression technologies and optimized storages are used, which provide high data processing speed and their movement between the components of the system.
Maia 100 supports high-speed Ethernet connections that allow you to process artificial intelligence models on a scale that are unattainable for traditional solutions. This technology provides reliable and balanced operation of the network, which is especially important for the tasks of confidential calculations.
Maia SDK software platform allows developers to quickly transfer their models written on Pytorch and Triton to new architecture, which significantly accelerates the process of deploying and optimizing models in Azure. SDK also includes tools for debugging, profiling and setting up models.
The system supports two software approaches, which provides flexibility and ease of development. Triton provides a convenient interface and automates memory management, while the Maia API offers a higher level of process control, which allows the maximum use of hardware capabilities.
Thus, Microsoft using Maia 100 significantly enhances its position in the field of cloud computing and artificial intelligence. The new architecture not only optimizes the work of AI models, but also reduces the costs of their deployment and operation, offering developers tools for more flexible and quick work.