RV1126是瑞芯微推出的一款编解码芯片,专门面向人工智能的机器视觉领域,支持4K编码,支持8路1080P同编同解,内置独立NPU,支持人脸识别、骨骼识别、分析预警等,可广泛应用于智慧安防、视频通讯、视频分析等场景,目前在智能安防摄像头、视频会议摄像头、DMS系统、行车记录仪等产品均有项目落地。
1.1 Overview
RV1126 is a high-performance vision processor SoC for IPC/CVR, especially for AI related application.It is based on quad-core ARM Cortex-A7 32-bit core which integrates NEON and FPU. There is a 32KB I-cache and 32KB D-cache for each core and 512KB unified L2 cache.The build-in NPU supports INT8/INT16 hybrid operation and computing power is up to 2.0TOPs. In addition, with its strong compatibility, network models based on a series of frameworks such as TensorFlow/MXNet/PyTorch/Caffe can be easily converted.
RV1126 also introduces a new generation totally hardware-based 14-megapixel ISP (image signal processor) and post processor. It implements a lot of algorithm accelerators usually used in IPC and CVR, such as HDR, 3A functions (AE, AF, AWB), LSC, 3DNR, 2DNR, sharpening, dehaze, fisheye correction, gamma correction, feature points detection and so on. All of them are real-time processing. Cooperating with two MIPI CSI (or LVDS/SubLVDS) and one DVP (BT.601/BT.656/BT.1120) interface, users can build a system that receives video data from 3 camera sensors simultaneous.
The video encoder embedded in RV1126 supports UHD H.265/H.264 encoding. It also supports multi-stream encoding, up to one 4Kp30 and one 1080p30 simultaneous. With the help of this feature, the video from camera can be encoded with higher resolution and stored in local memory and transferred another lower resolution video to cloud storage at the same time.The H.264/H.265 video decoder in RV1126 supports 4Kp30 for H.264 and H.265.
In addition to the previous high-performance multimedia block, RV1126 also contains rich audio, memory and other peripheral interfaces such as I2C, SPI, PWM and so on. These can help users add more sensors or other peripherals into whole system to improve flexibility and expansibility.
RV1126 has high-performance external DRAM (DDR3/DDR3L/DDR4/LPDDR3/LPDDR4-2133) capable of sustaining demanding memory bandwidths.
1.2 Features
The features listed below which may or may not be present in actual product, may be subject to the third-party licensing requirements. Please contact Rockchip for actual product feature configurations and licensing requirements.
1.2.1 Application Processor
Quad-Core Cortex-A7
Full implementation of the ARM architecture v7-A instruction set, ARM Neon Advanced SIMD
Separately Integrated Neon and FPU
32KB L1 I-Cache and 32KB L1 D-Cache per Cortex-A7 CPU
Unified 512KB L2 Cache for Quad-Core Cortex-A7
TrustZone technology supported
Separate power domains for CPU core system to support internal power switch and externally turn on/off based on different application scenario
PD_CPU0: 1st Cortex-A7 + Neon + FPU + L1 I/D Cache
PD_CPU1: 2nd Cortex-A7 + Neon + FPU + L1 I/D Cache
PD_CPU2: 3rd Cortex-A7 + Neon + FPU + L1 I/D Cache
PD_CPU3: 4th Cortex-A7 + Neon + FPU + L1 I/D Cache
One isolated voltage domain to support DVFS
1.2.2 Video Input Interface
Interface and video input processor
Two MIPI CSI/ LVDS/SubLVDS interfaces, 4 lanes each, MIPI CSI max data rate is 2.5Gbps/lane, LVDS/SublVDS max data rate is 1Gbps/lane
One 8/10/12/16-bit standard DVP interface, up to 150MHz input data
Support BT.601/BT.656 and BT.1120 VI interfaces
Support the polarity of pixel_clk、hsync、vsync configurable
RV1126 Datasheet Rev 1.4
ISP
Maximum resolution is 14Mpixel(4416×3312)
DVP input: ITU-R BT.601/656/1120 with raw8/raw10/raw12/raw16, YUV422
MIPI input: RX data lane x1/x2/x4, raw8/raw10/raw12, YUV422
3A: include AE/Histogram, AF, AWB statistics output
FPN: Fixed Pattern Noise removal
BLC: Black Level Correction
DPCC: Static/Dynamic defect pixel cluster correction
LSC: Lens shading correction
Bayer-NR: Bayer-raw De-noising, 2DNR
HDR: 3-/2-Frame Merge into High-Dynamic Range
TMO: 3-/2-Frame Merge Video Tone mapping
WDR: One Frame Wide-Dynamic Range Tone mapping
Debayer: Advanced Adaptive Demosaic with Chromatic Aberration Correction
CCM/CSM: Color correction matrix; RGB2YUV etc.
Gamma: Gamma out correction
Dehaze/Enhance: Automatic Dehaze and edge enhancement
3DLUT: 3D-Lut Color Palette for Customer
LDCH: Lens-distortion in the horizontal direction
Output Scale*3: support scale down level*3(W0 |