Specifications

CSD2-N96SPK3
Technical Specifications Server form

2U rack-mounted computing power server

Architecture

RISC-V architecture

Number of nodes

12 compute blades (96 distributed compute nodes) + 1 control node

Compute nodes

Octa-core 64-bit processor SpacemiT Key Stone K3, up to 2.4GHz

Control nodes

Octa-core 64-bit processor RK3588, main frequency up to 2.4GHz, the highest computing power is 6TOPS

AI computing power

5760TOPS (60T× 96, SquareINT4)

Fan module

14 high-speed cooling fans

Basic Specifications Video encoding
4K@90fps H.264
Video decoding

4K@180fps H.264/VP9

RAM

32GB LPDDR5 × 96
(8/16/32GB)

Storage

128GB UFS2.2 × 96

Storage expansion

3.5-inch/2.5-inch SATA3.0 SSD hard drive slot × 3 (Supports hot swapping; BMC can directly operate the hard drive, and computing child nodes can indirectly access the hard drive through the network sharing method provided by BMC)

Power

2 × 1300W hot-swappable power supplies, 1+1 redundancy support

Dimension

Standard 2U rack servers: 495.60mm × 928.51mm × 88.80mm

Environment

Operating Temperature: 0ºC ~ 30ºC, Storage Temperature: -40ºC ~ 60ºC, Operating Humidity: 5% ~ 80%RH (non-condensing)

Physical Specifications Installation requirements

IEC 297 Universal Cabinet Installation: 19 inches wide and 800 mm deep and above
Retractable slideway installation: The distance between the front and rear holes of the cabinet is 543.5mm~848.5mm

Software Specifications BMC

The BMC management system is integrated with the web-based management interface, supporting Redfish, VNC, NTP, monitoring advanced and virtual media, and the BMC management system can be redeveloped

Large language models

All models support private deployment of ultra-large-scale parameter models under the Transformer architecture, such as large language models including Deepseek-R1 Series, Gemma Series, Llama Series, ChatGLM Series, Qwen Series, Phi Series, etc.

Vision large models

K3: Supports private deployment of all vision large models
QCS8550: Supports private deployment of vision large models including Qwen2.5-VL, InternVL3, etc.

Deep learning

All models: Support traditional network architectures such as CNN, RNN, LSTM, and support various deep learning frameworks such as TensorFlow, PyTorch, PaddlePaddle, ONNX, and Caffe. Support custom operator development and Docker containerization management technology

Interface Specifications Internet

8 × 10Gbps SFP+, 1 × Gigabit Ethernet (RJ45, MGMT is used as BMC management network)

Console

1 × Console (RJ45, BMC debug serial port, baud rate 115200)

Video output

1 × VGA (maximum resolution 1080P, BMC management display)

USB

3 × USB3.0, 1 × Type-C (OTG)

Button

1 × Power, 1 × UID, 1 × Recovery, 1 × Reset

Customization

Firefly team provides hardware, software, turnkey and OEM/ODM customization services.