Maia 100
Microsoft's first in-house AI accelerator chip for Azure cloud.
The Maia 100 is a custom artificial intelligence accelerator designed by Microsoft for its Azure cloud infrastructure. Announced in November 2023 and built on a 5nm process by TSMC, it targets large-scale AI workloads like Copilot with high-bandwidth Ethernet networking and liquid cooling integration.
- Built on TSMC 5nm process with CoWoS-S packaging.
- Features 64GB HBM2E memory and 500MB cache.
- Uses custom Ethernet protocol with 4.8Tbps bandwidth.
- Designed for 500W production power usage.
- First chip to implement Microscaling (MX) data format.
What it is
The Maia 100 is a specialised processor created by Microsoft to handle artificial intelligence tasks in the cloud. It serves as the company’s first internally developed silicon for this purpose, aiming to optimise the entire stack from hardware to software. The chip is built to support complex models such as Microsoft Copilot, offering an alternative to third-party graphics processors for Azure services.
How it came about
Microsoft designed the Maia 100 based on its experience running large-scale AI operations. The development focused on co-designing hardware and software together to meet specific workload demands. The company announced the accelerator in November 2023, highlighting its role in enhancing Azure’s infrastructure. Further technical details were shared at the Hot Chips 2024 conference, revealing its architecture and performance targets.
Inside the chip
The chip features sixteen clusters, with each cluster containing four tiles. It includes image decoders and supports confidential computing. Memory consists of 64GB of HBM2E and a substantial 500MB combined L1 and L2 cache. The design supports various data types, including 9-bit and 6-bit compute formats. It implements the Microscaling (MX) data format to speed up training and inference. Networking relies on a custom Ethernet-based protocol rather than InfiniBand, providing 4.8 terabits per second of aggregate bandwidth.
How it was made
TSMC manufactures the Maia 100 using its 5nm process node and CoWoS-S advanced packaging technology. This approach allows for high density and performance. The physical design incorporates rack-level power distribution and closed-loop liquid cooling. A dedicated sidekick component matches the thermal profile of the main chip, helping to integrate the system into existing datacentre footprints while meeting zero waste commitments.
Where it went
This accelerator powers Azure AI infrastructure, specifically targeting cloud-based machine learning workloads. It runs models within frameworks like PyTorch and ONNX Runtime. Developers can use the Maia SDK or Triton programming language to deploy applications. The chip is intended for services such as Microsoft Copilot, where efficient scaling and lower operational costs are critical compared to using standard graphics processing units.
Versions and family
The Maia 100 stands as the initial implementation in the Azure Maia series. It introduces the v1.0 MX specification, developed in partnership with industry leaders like AMD, Intel, and NVIDIA. No direct predecessors or successors are listed in the provided material, marking it as the foundational unit for Microsoft’s custom silicon strategy in AI.
Why it matters
By creating its own silicon, Microsoft aims to reduce reliance on external suppliers and lower costs for AI computation. The chip promotes open standards through the Open Compute Project, encouraging broader adoption of the MX data format. Its Ethernet-centric design supports the Ultra Ethernet Consortium, potentially influencing how future data centres handle high-speed interconnects without proprietary networking hardware.
Problems and reception
Technical analysis notes the chip’s high power draw of up to 700W, though it operates at 500W in production. Observers suggest the reduced HBM capacity compared to competitors implies a focus on cost efficiency rather than raw peak performance. The use of HBM2E instead of newer memory types helps avoid supply chain competition. The integration with open-source tools like Triton is seen as a move to simplify developer adoption.
Questions people ask about Maia 100
Who manufactured the Maia 100 chip?
TSMC manufactured the Maia 100 using its 5nm process node and CoWoS-S advanced packaging technology. Microsoft designed the chip internally for its Azure cloud infrastructure.
What type of memory does the Maia 100 use?
The chip uses 64GB of HBM2E memory. It also includes a large 500MB combined L1 and L2 cache. This choice helps avoid competition for leading-edge high-bandwidth memory supplies.
How does the Maia 100 connect to other servers?
It uses a custom Ethernet-based network protocol instead of InfiniBand. Each accelerator provides an aggregate bandwidth of 4.8 terabits per second, supporting better scaling within Azure data centres.
What is the power consumption of the Maia 100?
The chip has a thermal design power of 700W. However, Microsoft states that it uses 500W per accelerator in production environments. It employs closed-loop liquid cooling for efficiency.
Good reading and sources
There is no English Wikipedia article on Maia 100 yet, so the facts on this page come from the sources listed above: azure.microsoft.com/en-us/blog/azure-maia-for-the-era-of-ai-from-silic, servethehome.com/microsoft-maia-100-ai-accelerator-for-azure/. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryChip wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.