Sophon BM1684 Chip: The 3rd-Gen AI Inference Powerhouse, Unleashing New Computing Potential for Full Cloud-Edge Scenarios

Location: Home News Updates Technical Articles

2026.05.20

4 次

BM1684

BM1684

As the AI inference industry continues to boom, edge computing and cloud deployment impose increasingly stringent requirements on chips. They need powerful computing performance, controlled power consumption and high compatibility to cater to diverse demands across all industries.

The third-generation AI inference chip BM1684 launched by Sophon (formerly Bitmain Sophon) precisely addresses these pain points. Boasting core strengths of high computing power, low power consumption and strong compatibility, it stands out as an optimal solution for cloud and edge AI inference scenarios.

As a core upgraded product of Sophon, the BM1684 delivers leapfrog performance improvements over its predecessor BM1682. It adds full INT8 computing support and achieves multiple-fold breakthroughs in key indicators such as computing power and video processing capabilities. It fully meets high-efficiency demands for deep learning inference, accelerating the transformation of AI applications from luxury facilities to essential infrastructure.

Powered by Robust Specifications, Balanced Computing Performance & Energy Efficiency

A chip’s true strength lies in its technical specifications. Adopting TSMC 12nm process technology, the BM1684 integrates powerful computing power within a compact size while maintaining optimal power consumption, achieving dual breakthroughs in high performance and low power consumption. Its core parameters are as follows:

✔ Maxed-out AI Computing PowerIt delivers up to 17.6 TOPS of INT8 computing power, which surges to 35.2 TOPS with Winograd acceleration, effortlessly handling high-density inference tasks. Its FP32 computing power reaches 2.2 TFLOPS, balancing precision and efficiency to adapt to AI models with varying accuracy requirements.

✔ Powerful Core ArchitectureEquipped with 8-core ARM Cortex-A53 processors running at a maximum main frequency of 2.3 GHz, together with 64 built-in NPUs (16 EUs per NPU), it features a total of 1024 EU computing units. It enables efficient and smooth computing scheduling to cope with complex inference scenarios with ease.

✔ Powerful Video Processing CapabilityIt supports decoding of 32-channel 1080P@30fps H.264/H.265 videos and enables real-time analysis of more than 10 high-definition video streams. It efficiently performs mainstream computer vision tasks such as face detection and license plate recognition, greatly improving video structured processing efficiency.

✔ Optimized Memory & Power ConsumptionCompatible with LPDDR4/LPDDR4X memory, expandable up to 16GB (6GB/12GB available on partial platforms) to meet high-bandwidth data processing demands. With a typical power consumption of only 16W, it boasts outstanding energy efficiency, perfectly fitting low-power operation requirements of edge devices without energy consumption concerns.

BM1684 >>>

Empowered by Full-stack Features, Boost Higher Efficiency in Development & Deployment

Beyond outstanding hardware specifications, the BM1684 also boasts comprehensive software support and rich interface expansion capabilities. It effectively lowers development barriers, improves deployment efficiency, and enables rapid implementation of computing power into practical applications.

Multi-framework Compatibility to Meet Diverse Development Needs

The chip is compatible with mainstream deep learning frameworks including TensorFlow, Caffe, PyTorch, MXNet, PaddlePaddle, ONNX and Darknet. It requires minimal algorithm migration and adaptation. Developers can select development tools according to their own habits, greatly shortening the development cycle and boosting R&D efficiency.

Complete Toolchain for Worry-free One-stop Development

It is equipped with the all-in-one BMNNSDK2 development toolkit, including compilers, inference engines, quantization tools and Docker container support. It supports containerized deployment and K8s scheduling. Featuring mature, stable and user-friendly toolchains, it enables both novice developers and senior engineers to get started quickly and achieve efficient development and deployment.

Rich Interfaces for Easy Integration into Various Devices

It is equipped with abundant interfaces including dual Gigabit Ethernet, USB 3.0/2.0, HDMI, mSATA, PCIe, RS485/RS232, featuring excellent compatibility. It can be easily embedded into various edge devices and cloud servers without extra interface adaptation, lowering equipment integration difficulty and accelerating product launch.

Full-scenario Coverage, Empowering Intelligent Upgrade Across All Industries

Leveraging its core advantages of high computing power, low power consumption and easy deployment, BM1684 has been widely applied in smart security, smart transportation, smart retail and many other fields, serving as a core computing pillar driving industrial intelligent upgrading. In addition, it can adapt to products with diverse computing power requirements via computing power stacking, covering full-scenario applications across cloud, edge and terminal ends.

Smart Security It supports facial recognition, behavior analysis and other functions, and can be applied to smart cameras, security terminals and other devices to realize all-weather real-time monitoring and abnormal early warning, thus improving security management efficiency.
Smart Transportation It efficiently fulfills license plate recognition, vehicle-road collaboration and other tasks, and adapts to scenarios such as traffic monitoring and intelligent checkpoints. It facilitates intelligent traffic management and eases traffic pressure.
Smart Retail Widely used in unmanned supermarkets, passenger flow statistics and other scenarios, it accurately collects passenger flow data and analyzes consumption behaviors, providing decision-making basis for retailers and boosting operational efficiency.
Other Scenarios It delivers outstanding performance in industrial quality inspection, drones, smart classrooms, video structuring and more. It provides stable and efficient computing support for intelligent transformation of various industries. It also empowers educational scenarios and accelerates the implementation and application of large AI models on educational terminals.

Empowered by Computing Power, Embark on a New Journey of Cloud-edge Inference

Nowadays, edge AI inference is gaining increasing popularity. Built on TSMC's 12nm process technology, the BM1684 chip strikes a perfect balance between high computing power and low power consumption. Equipped with a complete toolchain, abundant interfaces and wide scenario adaptability, it stands as the core pillar of Sophon's full-scenario computing product lineup and serves as vital support for the implementation of AI applications across diverse industries.

Whether for high-density inference on the cloud or low-power deployment at the edge, the BM1684 can precisely meet diverse demands. It provides developers with efficient and convenient computing solutions, helping AI technologies move from laboratories to real-world scenarios and accelerating the intelligent upgrading of all industries.

In the future, as Sophon’s ecosystem continues to improve, the BM1684 will keep giving full play to its computing advantages. Joining hands with more partners, it will explore more possibilities for AI applications and make high-efficiency computing power benefit all industries.