- Joined
- May 21, 2024
- Messages
- 506 (3.64/day)
Microsoft provided a detailed view of Maia 100 at Hot Chips 2024, their initial specialized AI chip. This new system is designed to work seamlessly from start to finish, with the goal of improving performance and reducing expenses. It includes specially made server boards, unique racks, and a software system focused on increasing the effectiveness and strength of sophisticated AI services, such as Azure OpenAI. Microsoft introduced Maia at Ignite 2023, sharing that they had created their own AI accelerator chip. More information was provided earlier this year at the Build developer event. The Maia 100 is one of the biggest processors made using TSMC's 5 nm technology, designed for handling extensive AI tasks on Azure platform.
Maia 100 SoC architecture features:
The Maia 100 uses a network connection based on Ethernet with a special protocol similar to RoCE, which allows for very fast data processing. It can handle up to 4800 Gbps for certain data operations and 1200 Gbps for all-to-all communication.
Maia 100 Specs:
The Maia Software Development Kit (SDK) helps programmers adapt their PyTorch and Triton models for use with Maia. The SDK includes various tools to make it simple to use these models with Azure OpenAI Services. Programmers can write code for the Maia system using either Triton, an open source domain specific language (DSL) for deep neural networks, or the Maia API, a custom model that offers high performance and detailed control. Maia directly supports PyTorch, so developers can run PyTorch models with only minor adjustments to their code.
It'll be interesting to see if Microsoft decides to open up access to Maia 100 accelerators to other organizations, similar to what Google and Amazon have done with their AI chips.
View at TechPowerUp Main Site | Source
Maia 100 SoC architecture features:
- A high-speed tensor unit (16xRx16) offers rapid processing for training and inferencing while supporting a wide range of data types, including low precision data types such as the MX data format, first introduced by Microsoft through the MX Consortium in 2023.
- The vector processor is a loosely coupled superscalar engine built with custom instruction set architecture (ISA) to support a wide range of data types, including FP32 and BF16.
- A Direct Memory Access (DMA) engine supports different tensor sharding schemes.
- Hardware semaphores enable asynchronous programming on the Maia system.
The Maia 100 uses a network connection based on Ethernet with a special protocol similar to RoCE, which allows for very fast data processing. It can handle up to 4800 Gbps for certain data operations and 1200 Gbps for all-to-all communication.
Maia 100 Specs:
- Chip Size: 820 mm²
- Design to TDP: 700 W
- Provision TDP: 500 W
- Packaging: TSMC N5 process with COWOS-S interposer technology
- HBM BW/Cap: 1.8 TB/s @ 64 GB HBM2E
- Peak Dense Tensor POPS: 6bit: 3, 9bit: 1.5, BF16: 0.8
- L1/L2: 500 MB
- Backend Network BW: 600 GB/s (12X400gbe)
- Host BW (PCIe): 32 GB/s PCIe Gen5X8
The Maia Software Development Kit (SDK) helps programmers adapt their PyTorch and Triton models for use with Maia. The SDK includes various tools to make it simple to use these models with Azure OpenAI Services. Programmers can write code for the Maia system using either Triton, an open source domain specific language (DSL) for deep neural networks, or the Maia API, a custom model that offers high performance and detailed control. Maia directly supports PyTorch, so developers can run PyTorch models with only minor adjustments to their code.
It'll be interesting to see if Microsoft decides to open up access to Maia 100 accelerators to other organizations, similar to what Google and Amazon have done with their AI chips.
View at TechPowerUp Main Site | Source
and microsoft tell us ...we have a new AI chip