Maia 300

Maia 300 uses a multi-chiplet design: the compute chiplet is based on TSMC's N2P process, while the memory controllers and I/O functions are placed on separate dies.

Microsoft claims Maia 300 should deliver 40% more tokens per megawatt than leading GPU systems, reaching 7 times the rack-level token throughput of Maia 200 in a 96-accelerator configuration.

Maia 300 accelerators can connect to remote peers in the superpod at speeds of 750GB/s.

Maia 300's scale-up network runs a custom AI transport over Ethernet, with a two-tier topology: an intra-rack half-switch architecture electrically connects all 96 accelerators, and a second optical tier connects up to 128 racks into a single superpod (12,288 accelerators).

Liquid cooling handles more than 80% of the Maia 300 rack's thermal load, and the design can connect to facility water or adapt to air-cooled data centers via heat exchanger units.

SiliconANGLE's analysis holds that Maia 300's 288GB capacity matches NVIDIA's Rubin accelerator, but its memory bandwidth is about 20% lower.

Maia 300 is configured with 8 HBM4 stacks, providing 288GB of capacity and 18TB/s of bandwidth, which is 2.5 times the memory bandwidth and 33% more capacity than Maia 200.

Maia 300 is claimed to support multi-gigawatt-scale inference deployments on Azure and to become the primary inference system for Microsoft's own products.

Maia 300 is equipped with 8 HBM4 memory stacks, uses compute chiplets manufactured on TSMC's N2P process, and is deployed in the form of 96-accelerator racks.

The key metric Microsoft gives is that Maia 300 improves token throughput per megawatt of power consumption by 40% compared with the previous generation.

Microsoft's advancing Maia while also purchasing Nvidia GPUs is essentially a combined configuration between generality and specialized efficiency, rather than a complete replacement.

Through its self-developed Maia 300, Microsoft aims to optimize the execution efficiency of large language models and complex reasoning models and to reduce its dependence on third-party GPU suppliers.

Maia 300 was tuned for the Azure cloud environment from the outset of its design.

The 40% efficiency improvement is an official figure based on specific benchmarks, and actual performance is affected by factors such as model structure, batch size, memory bandwidth and interconnect bandwidth, so it should be viewed as a comparison under ideal conditions.

Maia 300 falls into the ASIC (application-specific integrated circuit) category, in contrast to the general-purpose programmable architecture of Nvidia GPUs.

Maia 300是Maia系列的迭代产品，命名上跳过了Maia 200，官方尚未给出明确说明。

One of Maia 300's key upgrades is greater capability to handle the latest inference models and agentic tasks.

One of Maia 300's goals is to reduce Microsoft's dependence on Nvidia GPUs while improving the cost and efficiency of cloud AI workloads.

Production of Maia 300 will start in 2027.

Microsoft plans to have Maia 300 gradually replace the third-party chips that Microsoft's internal inference workloads rely on.