Software-Hardware Integration to Accelerate
Software-hardware integration: the optimal path to mass production efficiency
Horizon Robotics was among the first in the industry to propose—and continuously practice—a software-hardware integration approach to smart driving. To truly deliver co-optimization, it takes not only world-class algorithms and software, but also purpose-built computing architecture—and the ability to unlock ultimate system efficiency in mass-production engineering. This is why Horizon Robotics can break through the constraints of traditional programming paradigms and Moore’s Law, and continue to lead the upgrade of automotive intelligence.
-
2015
Horizon Robotics was founded as the world’s first company dedicates to specialized computing solutions for deep neural networks
-
2016
Horizon Robotics introduced China’s first self-developed embedded neural network processor—BPU™
-
2017
Horizon Robotics launched China’s first automotive computing solution, Journey, pioneering large-scale pre-installed mass production
-
2019
Horizon Robotics released the TogetheROS™ in-vehicle operating system, providing a safe and real-time development kit for production-grade assisted driving
-
2021
Horizon Robotics accelerated its software-hardware integration ecosystem, working with partners to build a flexible, mature open platform for production-grade assisted driving
-
2024
Horizon Robotics launched Horizon SuperDrive (HSD™), a full-scenario assisted driving solution integrating Horizon's software and hardware technologies—resetting the benchmark for urban driving assistance
-
2015
Horizon Robotics was founded as the world’s first company dedicates to specialized computing solutions for deep neural networks
-
2016
Horizon Robotics introduced China’s first self-developed embedded neural network processor—BPU™
-
2017
Horizon Robotics launched China’s first automotive computing solution, Journey™, pioneering large-scale pre-installed mass production
-
2019
Horizon Robotics released the TogetheROS™ in-vehicle operating system, providing a safe and real-time development kit for production-grade assisted driving
-
2021
Horizon Robotics accelerated its software-hardware integration ecosystem, working with partners to build a flexible, mature open platform for production-grade assisted driving
-
2024
Horizon Robotics launched Horizon SuperDrive (HSD™), a full-scenario assisted driving solution integrating Horizon's software and hardware technologies—resetting the benchmark for urban driving assistance
Driving paradigm-level
innovation in algorithms
Founded by leaders in computer vision and machine learning, Horizon Robotics has driven the development of deep learning across industry, academia, research, and applications. With world-class algorithm capabilities, Horizon Robotics is a pioneer in deploying advanced smart driving approaches such as end-to-end and the use of interactive game theory, and will continue to advance the software-defined vehicle transformation.
-
World ChampionHorizon Robotics has won five world titles in the Google Waymo Open Dataset Challenge
-
No.1 on the LeaderboardSparse4D—Horizon's end-to-end perception algorithm—ranks No.1 on the nuScenes public dataset
-
Best PaperUniAD, a large-model approach to end-to-end smart driving proposed by Horizon researchers, won the CVPR 2023 Best Paper Award
The Optimal Choice for Smart Computing
Horizon Robotics' proprietary smart computing architecture, BPU™, focuses on the latest advances in neural network architectures and intelligent accelerated computing. It is optimized specifically for smart driving use scenarios to deliver true automotive efficiency. The Journey™ family consistently integrates the BPU intelligent acceleration unit, fully leveraging the high-performance, low-latency, and low-power advantages of computing solutions to maintain generation-leading competitiveness.
Hardware as the Engine
Accelerating intelligent evolution
After three generations of architectural iteration, BPU™ has proven itself in large-scale mass-production deployment and will further open up through IP licensing to support OEMs’ in-house smart driving computing solution development. By integrating algorithms, compilers, and computing architectures, BPU can evolve through data-driven automation and verification, continuously delivering the optimal compute-efficiency solution for smart driving systems and powering faster evolution of smart vehicles.
Purpose-built for end-to-end, full-scenario assisted driving
Binarized Probability Distribution
For ADAS Scenarios
Object Detection, Semantic Segmentation
Generative Model
For Highway NOA
2.5D/3D Vision Algorithms, Object Tracking, Trajectory Prediction
Game Theory
For Full-scenario NOA
Temporal Environment Prediction, Interactive Rules in Complex Environments
Purpose-built for end-to-end, full-scenario assisted driving
Horizon Robotics has introduced a new intelligent acceleration engine built for next-generation smart driving systems — BPU™ Nash. Featuring a highly heterogeneous compute core, Nash increases compute diversity and integrates tightly with the compiler for automatic optimization, greatly improving programmability. With extreme optimization for frontier algorithms, Nash delivers outstanding performance well beyond its class, achieving the best algorithm efficiency.
Learn More
-
三级存储架构-
极致优化大参数下的带宽瓶颈
-
核间高效协同
-
-
数据变换引擎-
灵活支持Transformer细小算子
-
-
多脉动立方加速引擎-
灵活的引擎间数据流动
-
高能效、低宽带占用
-
-
浮点向量加速单元-
通用、灵活
-
关键算子精度需求
-
-
紧耦合异构计算-
高效加速不同类型数据处理
-
-
多向数据流动-
核内、核间、片间高效灵活的数据流动
-
动态调度，灵活调优
-
-
虚拟化 Virtualization-
透明式提升多任务并行处理能力
-
-
数据驱动功耗优化-
针对神经网络数据动态范围特性
-
降低功耗 30%
-
高效支持大参数Transformer
-
强大的并行浮点算力-
支持线程并发的 SIMT Vector Processing Unit (VPU)
-
支持 BF16/FP16/FP32 多种浮点数据类型，在性能和精度之间取得更好的平衡
-
-
特别优化的超越函数-
支持 Layer-norm & Softmax 算子的硬件加速
-
支持 Transpose & Reshape 算子的硬件加速
-
-
全新的存储系统-
L0M, L1M, L2M 三级存储系统用于数据缓冲和交换
-
先进总线架构配合高带宽 DDR，有效缓解内存墙问题
-
原生支持大规模交互式博弈
-
算力灵活可编程支持用户自定义算法 -
灵活满足博弈算法完整支持矩阵、向量、标量计算和逻辑控制 -
灵活处理各类任务专门为高频次小模型调度计算设计的标量计算核（RISCV）和 Vector/Tensor 计算核紧耦合交互
Pursuing Real Computing Performance
A New Moore's Law for Smart Computing
In the era of smart computing, theoretical peak compute (TOPS) does not reflect real performance, and does not directly translate into consumer experience. As early as 2016, Horizon Robotics proposed a new Moore’s law based on real computing performance, arguing that FPS (Frames Per Second) better captures the real efficiency of advanced deep learning algorithms running on hardwares.
FPS/$
Real Computing Performance
TOPS/$
Theoretical Peak
Compute Efficiency
Utilization
Effective Utilization
FPS/TOPS
Algorithm Efficiency
Big Compute Matters. Faster Compute Wins.
Software-hardware integration has become an industry consensus as the key path to breaking performance bottlenecks. From Bernoulli to Nash, BPU™ has successfully gone beyond the limits of Moore’s Law. Turning complexity into simplicity, Horizon Robotics converges simplified algorithms, end-cloud integrated data, and co-optimized computing into systematic technical strength — pushing computing performance to continuously raise the ceiling of consumer experience.
Moore's Law
New Moore's Law
* 1.ResNet-18 720P 2.Vit-base-patch16 224x224
Horizon Robotics remains committed to its founding principle of "People First, Technology for People". We open our full-stack software–hardware integration capabilities to the industry, to enhance developer productivity and make mass-production deployment more efficient and achievable.
To further support talent development in smart driving, Horizon Robotics offers developer mentoring, training courses, research and study collaborations, entrepreneurship support, and open skills certifications—helping developers continuously improve and build hands-on experience.
Together, Expanding the Boundaries of Innovation
