CSD2-N96 – 96-Node AI Compute Server 2U Rack

96-Node AI Compute Server 2U Rack

Maker, Prototyping, Open-Hardware & Developer Boards

Manufacturer part number
CSD2-N96
Manufacturer
Firefly

The CSD2-N96 is a 2U rack AI compute server with 12 compute blades hosting 96 nodes, designed for private deployment of large language models and vision large models.

Product overview

Overview. The Firefly CSD2-N96 is a 2U rack AI compute server with 12 compute blades, providing 96 distributed compute nodes plus one control node. It is offered in several processor configurations, including QCS8550 (16G+256G), RK3588S (16G+256G), RK3576 (8G+64G), BM1688 (8G+64G) and SPK3 (32G+128G).

Computing power. The control node is an octa-core 64-bit RK3588 processor with a main frequency up to 2.4 GHz and a maximum computing power of 6 TOPS. Total AI computing power is 4608 TOPS (48T × 96, INT8) in the 576 TOPS (6T × 96, INT8) configuration and 5760 TOPS (60T × 96, SquareINT4) at the highest level. The architecture is ARM or RISC-V depending on the configuration.

Interfaces and connectivity. The server provides 8 x 10 Gbps SFP+ ports and one Gigabit Ethernet RJ45 port used as the BMC management network. It also includes 3 USB 3.0 ports, one Type-C (OTG) port, a VGA output up to 1080P for BMC management display, and an RJ45 console port at 115200 baud for BMC debugging.

Power and environment. The server is powered by 2 hot-swappable 1300W power supplies with 1+1 redundancy and cooled by 14 high-speed fans. It operates between 0°C and 30°C at 5% to 80% RH non-condensing and can be stored between -40°C and 60°C.

Storage. Three SATA 3.0 slots for 3.5-inch or 2.5-inch drives are hot-swappable. The BMC can operate the drives directly, and compute nodes access them through network sharing provided by the BMC.

Software and management. The BMC management system integrates a web interface and supports Redfish, VNC, NTP and virtual media, and it can be redeveloped. The server supports private deployment of language models such as the Llama, Qwen, Gemma and Phi series, and vision models such as Qwen2.5-VL or InternVL3 depending on the platform. It supports TensorFlow, PyTorch, PaddlePaddle, ONNX and Caffe frameworks, as well as Docker containerization.

Integration. The chassis mounts in a standard IEC 297 19-inch cabinet at least 800 mm deep, or on retractable slideways for a front-to-rear hole distance of 543.5 mm to 848.5 mm. Dimensions are 495.60 mm × 928.51 mm × 88.80 mm.

See also

Technical questions

Ask a question

No published question about this part yet. Questions are reviewed before publication.

Request a quote