Article Directory
- Application : HUAWEI Da Vinci Architecture
https://support.huaweicloud.com/intl/en-us/odevg-A800_9000_9010/atlaste_10_0007.html - PPT : DaVinci: A Scalable Architecture for Neural Network Computing
https://www.cmc.ca/wp-content/uploads/2020/03/Zhan-Xu-Huawei.pdf
related blog
- In-depth interpretation of the Da Vinci architecture: the "secret weapon" of Huawei's AI chip
https://baijiahao.baidu.com/s?id=1642553630954055619 - Ascend AI Processor: Da Vinci Architecture
http://sfyh.hnbdata.cn/aipath/frames/ascend/intro/davinci.html - Wu Jianming wujianming: Da Vinci Architecture NPU
https://zhuanlan.zhihu.com/p/352481421
- computing unit
- Cube
- Vector
- Scalar
- storage unit
- control unit
As the computing core of the Ascend AI processor, the AI core is responsible for executing computation-intensive vector and tensor operators.
It can be seen as a simplified structure of a modern microprocessor. It includes three basic computing resources: matrix computing unit (cubic unit), vector computing unit (vector unit) and scalar computing unit (scalar unit).
These three computing units correspond to three common computing modes: tensor, vector, and scalar.
In the actual computing process, the three computing units each perform their duties, forming three independent execution pipelines, which cooperate with each other under the unified scheduling of the system software to achieve the optimization of computing efficiency.
In addition, cube cells and vector cells support different precision and different types of calculation modes.
AI Core architecture
The AI core is composed of computing unit, storage unit and control unit.
- Computing units
The execution units in the AI core include Cube, Vector, and Scalar, which are computing units for different types of data. - The storage unit
AI Core loads external data into the internal memory for calculation.
Programmer-visible internal storage devices include L1 buffer, L0 buffer, unified buffer, general-purpose registers (GPR), special-purpose registers (SPR), and scalar buffers.
To facilitate data transfer and movement in the AI Core, a Bus Interface Unit (BIU), Memory Transfer Engine 1 (MTE1), MTE2, and MTE3 are provided.
BIU provides an interface for the interaction between the AI core and the bus.
MTE moves data between different buffers. - Control Unit
The control unit of AI Core includes System Control, Scalar PSQ, Instr.Dispatch, Cube Queue, Vector Queue, MTE Queue and Event Synchronization. System Control is responsible for the operating mode, parameter configuration and power consumption control of AI Core.
Scalar PSQ is mainly used to decode control instructions.
After the instructions are decoded and sequentially transmitted using the Instr.Dispatch module, they are sent to the Cube Queue, Vector Queue or MTE Queue module respectively by type.
Huawei AI panorama
Iori 2023-08-04 (Friday)