LIVETracking 971 AI models from 268 providers.
Beta
Microsoft logoMicrosoft

Microsoft Maia 100 AI Accelerator Rack

Full rackRampingMaia 1002025

Microsoft's first custom AI silicon at rack scale. Maia 100 chips fabricated at TSMC on N5 with HBM3e. Designed for Azure AI inference and fine-tuning. Co-designed with the Cobalt ARM CPU for host compute. Microsoft's strategic hedge against NVIDIA dependency.

32
GPUs per system
96 FP8 PFLOPS
All systemsCompute hub
GPU model
MM
Microsoft Maia 100
GPU count
32x
CPU model
Microsoft Cobalt 100
8x
Memory type
H
HBM3e
Total HBM
6.1 TB
Host memory
8 TB
Interconnect
Custom Ethernet fabric
8 TB/s
Networking
1600 Gbps
Storage
60 TB NVMe SSD
Form factor
Custom rack
Weight
TBD
Rack units
42U

Manufacturer datasheet values · aggregate system compute

FP4 PFLOPSTBD
FP8 PFLOPS96
FP16 PFLOPS48
BF16 PFLOPS48
Training effective PFLOPS36

Thermal envelope · cooling requirements · efficiency

Rack power
40 kW
Per GPU
500 W
Cooling
liquid
PUE estimate
1.1
Power draw relative to tracked systems40 kW / 2500 kW max
2.4 FP8 PFLOPS per kW · average across all systems is 4.81

Hardware amortized over 3 years · power at $0.05/kWh

List price
TBD
PFLOPS per kW
2.4
Available from
MA
Microsoft Azure

Disclosed in press releases, SEC filings, and conference talks

Microsoft Azure logoMicrosoft Azure
Quantity
Internal deployment
Source
Microsoft Ignite 2024

Every data point on this page is reproducible

Compare across the system landscape

Google logoTPU v5p Pod
8960x Google TPU v5p · 8100 PFLOPS
Pod / clusterShipping
NVIDIA logoDGX GB300 NVL72
72x NVIDIA B300 · 1440 PFLOPS
Full rackAnnounced
NVIDIA logoDGX GB200 NVL72
72x NVIDIA B200 · 720 PFLOPS
Full rackShipping
Google logoTPU v6e Pod
256x Google TPU v6e · 230 PFLOPS
Pod / clusterShipping
NVIDIA logoDGX B200
8x NVIDIA B200 · 80 PFLOPS
Server nodeShipping
AMD logoMI325X Platform
8x AMD MI325X · 48 PFLOPS
Server nodeShipping
Back to all systemsCompute hub