Firefly launched two array server products, both adopting a 1+10 distributed array architecture. One RK3588 chip serves as the BMC management unit, flexibly supporting the expansion of 0 to 10 computing power core boards for elastic computing power scaling. The core advantage of the device is that it can be upgraded with the Houmo AI M50 computing power module, enabling an upgrade from conventional small AI model inference to the local deployment of 30B large models, precisely matching the computing power iteration needs of government and enterprise AI projects.
Both servers have a built-in self-developed BMC system, supporting remote management, status monitoring, and batch firmware management. This makes operation and maintenance simple and efficient, significantly reducing later-stage O&M and upgrade costs.
Hardware Parameter Comparison of the Two Models
Core Business Capabilities
01. Video Encoding and Super-Storage Compression Scenarios
Leveraging the native 8K hardware encoding/decoding capability of the RK3588, it can perform secondary intelligent compression on H.264 and H.265 HD videos, achieving up to 10x storage compression. This effectively reduces storage space, bandwidth, and cloud storage costs, making it suitable for scenarios such as security video archiving, streaming media backhaul, and massive video storage.
02. AI Algorithm Inference Scenarios
The device is natively compatible with the full series of YOLOV algorithms and can stably handle AI inference tasks such as image recognition, behavior analysis, passenger flow statistics, and voice translation. To meet the needs of standardized project implementation, it provides standardized video AI computing power configurations of 64 / 96 / 192 / 256 channels. The node ratio is precise, preventing computing power waste or insufficient performance. Compared with traditional X86 server clusters, it has significant advantages in cost-effectiveness and implementation adaptability.
Core Board Selection Guide: The entire machine supports targeted matching of business needs and differentiated adaptation of computing power core boards.
03. 600-Day Data Ultra-Storage
The CSB2-N10 is equipped with 6 bays for 3.5-inch / 2.5-inch hard drives, supporting SATA 3.0 HDD / SSD expansion for easy storage scaling. It has a built-in data redundancy protection mechanism, ensuring that business is not interrupted and data is not lost in case of a single disk failure, significantly improving system reliability. At the same time, it reserves ample expansion space, making forward-looking preparations for future business growth.
04. 30B Large Model AI Computing Power Upgrade
Based on the original edge AI server, this solution is equipped with the Houmo Manjie M50 Processing-in-Memory (PIM) acceleration module. Based on a domestic PIM architecture, it achieves low-power, high-performance edge computing power upgrades. The core advantages are as follows:
High Cost-Performance Inference Capability
Supports the smooth local operation of 30B-level open-source large models, completing computing power leapfrogging with a lower TCO, breaking the limitation of traditional edge devices that can only handle shallow inference.
Seamless Compatibility with Existing Facilities
No need to replace the entire machine or adapt the software ecosystem. Computing power can be upgraded simply by adding an AI computing card, lowering the deployment threshold.
Flexible and Scalable Computing Architecture
Supports on-demand allocation and dynamic deployment, adapts to hybrid deployment scenarios of large and small models, and reserves elastic space for future computing power upgrades.
Performance Parameters and Test Data of the Houmo M50 Computing Module
Tested Inference Performance of Mainstream Large Models with Houmo M50
To review the full series of 8B / 20B / 30B models, throughput data for different context lengths and multi-batch loads, and the complete official performance benchmark report, please visit the Houmo Developer Official Documentation.
Typical Scenario: AI Teaching Assistant
Real-time classroom audio is automatically transcribed into standard text. The edge-side 30B large model then completes the extraction of core knowledge points and structured content sorting, intelligently generating standard classroom notes and teaching minutes. The entire process runs offline on an intranet, ensuring security and compliance. Teachers can say goodbye to tedious post-class organization work and truly devote their energy to teaching optimization and student interaction, significantly improving review efficiency.
Experience the CSB2-N10 Server
The CSB2-N10 is now officially available. Friends with edge-side 30B large model application scenario needs are welcome to contact our business consultant: sales@t-firefly.com
