CSA1-N8S1684 Overview Specifications Wiki Technical Case Download

CSA1-N8S1684 AI Server

Powered by the third-generation TPU BM1684, this server is equipped with up to 8 BM1684 computing modules. It supports up to 140.8TOPS
of INT8 peak computing power or 17.6TOPS of FP32 high-precision computing power, with the power of multi-channel video encoding and
decoding. It provides multiple algorithm migration and a one-stop deep learning development toolkit. With the BMC management system, it
supports secondary development.

BM1684 octa-core AI SoC
with high computing power

Powered by BM1684 SoC with AI computing power, each computing module features octa-core Cortex-A53 architecture, 2.3GHz
frequency and up to 17.6Tops of AI computing power. It provides 12GB LPDDR4 and 128GB eMMC large storage capacity.

Eight computing modules, flexible configuration

The server can be equipped with up to 8 BM1684 computing modules. Users can customize the number and storage configuration of
computing modules, which are flexibly used in various scenarios.

140.8TOPS powerful computing power

It supports up to 140.8TOPS of INT8 peak computing power or 17.6TOPS of FP32 high-precision computing power, which can
meet the application requirements of deep-learning model development and training.

Incredible multi-channel video processing capability

With the power of multi-channel video processing, it supports up to 128-channel 1080P HD video whole-process (decoding + AI)
processing, which can bring excellent performance.

Various algorithms, strong adaptability

This server supports algorithm migration of pedestrian/vehicle/object recognition, video structuring, trajectory behavior
analysis, etc, featuring high security and reliability. So it can be flexibly applied to various product development.

Open SDK, one-stop AI development kit

SOPHON SDK (BMNNSDK2), one-stop deep learning development toolkit provides a series of software tools including the underlying
driver environment, compiler and inference deployment tool. It supports mainstream frameworks: Caffe/TF/PyTorch/Mxnet/Paddle, mainstream
network model and custom operator development, Docker containerization, and rapid deployment of algorithm applications.

BMC management system

BMC intelligent management system provides a visualization console interface through HDMI output. It easily achieves real-time monitoring,
software configuration, hardware management, troubleshooting and system upgrade. The system supports secondary development.

Dual 10GbE meet high bandwidth requirements

Dual 10GbE ports and a GbE port meet business scenarios with high bandwidth requirements. The independent BMC management network
interface separates management network and service network, ensuring the security and reliability of network communication.

Support SATA3.0 hard disk expansion

A 3.5-inch hard disk bay supports SATA3.0 HDD/SSD hard drive expansion, allowing the device to be
easily expanded to TB storage capacity.

Standard 1U rack server

Standard 1U rack server chassis design, its depth is 576mm, fully matching most types of cabinets in the data center.

A wide range of application scenarios

It supports flexible deployment in multiple fields and scenarios, such as intelligent park, security surveillance, intelligent retail,
intelligent transportation, intelligent finance, intelligent city, intelligent industry, smart energy and more.

Intelligent park

Security surveillance

Intelligent retail

Transportation

Intelligent finance

Intelligent industry

Specifications

Specifications
Product Name

Computing Power Cluster Server
Product Model

CSA1-N8S1684
AI Computing Power

140T@INT8 peak computing power, 17.6TFLOPS@FP32 high-precision computing power
VPU

Video decoding: H.264 & H.265: 1080P @11520fps

Video decoding resolution: 8192 * 8192 / 8K / 4K / 1080P / 720P / D1 / CIF

Video encoding: H.264 & H.265: 1080P @600fps

Video encoding resolution: 4K / 1080P / 720P / D1 / CIF

JPEG decoding: JPEG: 4800 images per second @1080P
Nodes

8 compute nodes (up to 64 ARM cores) + 1 control node
CPU

Compute node: BM1684 octa-core (A53*8) 64-bit processor, frequency up to 2.3GHz

Control node: RK3588S octa-core (4×Cortex-A76+4×Cortex-A55) 64-bit processor, frequency up to 2.4GHz
RAM

12GB LPDDR4 × 8 (compute nodes)
Storage

32GB eMMC × 8 (compute nodes)

3.5-inch SATA3.0 hard disk × 1
Network

GE  RJ45 Gigabit Ethernet port×2 (a control node port and a common Ethernet port)

SFP 10GE × 2

4G/LTE/5G network (optional)
Display

HDMI 2.0, 4K@60Hz (main processor core board display)
USB

2 × USB3.0 HOST, 1×Type-C (for processor core board debugging)
Fan

5 high-speed fans
Power

300W AC power supply (input: 100V AC～240V AC)
OS

Linux
BMC

Integrated BMC management system provides management interface based on Web.

This system supports secondary development.
Deep Learning Framework

TensorFlow / PyTorch / Paddle / Caffe / ONNX / MXNet / DarkNet
Dimension

Standard 1U rack-mounted server: 490mm x 390mm x 44.4mm
Operating Temperature

0ºC - 50ºC
Operating Humidity

8%RH~95%RH