3. Pulsar Toolchain Overview

3.1. Introduction

Pulsar is an all-in-one neural network compiler developed by Irisoft self-developed, It is a quadruple-play of conversion, quantization, compilation, and heterogeneous to achieve fast and efficient deployment of deep learning neural network models. We have optimized the features of AX6 series chips to fully utilize the computing power of on-chip heterogeneous computing units (CPU+NPU) and improve the efficiency of neural network model deployment.

The core function of the Pulsar toolchain is to compile .onnx models into .joint models that can be parsed and run by the chip.

Architectural diagram

Chip

Deployment Process

pipeline

3.2. Virtual NPU Introduction

AXera currently provides AX630A, AX620A, and AX620U available chips, and the following is a brief description of the virtual NPU hardware configuration information for these chips.

3.2.1. AX630A

Virtual NPU (Neural-network Processing Unit)

Mode

Support List

DEFAULT

0

2_2

221、222

3_1

311、312

Note

Only one mode can be selected for NPU hardware initialization
DEFAULT mode: no division of NPU resources, use all NPU resources
2_2 mode: divide NPU resources in half, use half of NPU resources for 221 and 222 each
3_1 mode: unequal division of NPU resources, 311 mode has roughly three times as many resources as 312
Virtual NPU mode, you can run PTQ and AI-ISP models at the same time

3.2.2. AX620A

Virtual NPU

Mode

Support List

DEFAULT

0

1_1

111、112

Note

Only one mode can be selected for NPU hardware initialization
DEFAULT mode: no division of NPU resources, use all NPU resources
1_1 mode: divide NPU resources in half, 111 and 112 each use half of NPU resources
With virtual NPU mode, you can run PTQ and AI-ISP models simultaneously

3.2.3. AX620U

The AX620U is the same as the AX620A in terms of the above functions.