Firefly Technology, according to the CNX-Software resource, has released CSB1-N10 series servers for building AI clusters. Devices in a 1U form factor are suitable for data centers, edge deployments and local sites.
The family includes five models: CSB1-N10S1688, CSB1-N10R3588, CSB1-N10R3576, CSB1-N10NOrinNano and CSB1-N10NOrinNX. The system nodes are respectively equipped with Sophgo Sophon BM1688 processor (8 cores, up to 1.6 GHz), Rockchip RK3588 (8 cores, up to 2.4 GHz), Rockchip RK3576 (8 cores, up to 2.2 GHz), NVIDIA Jetson Orin Nano ( 6 cores, up to 1.5 GHz) and Jetson Orin NX (8 cores, up to 2.0 GHz).
The first three of the listed nodes are equipped with LPDDR4 RAM with a capacity of 8, 16 and 8 GB, respectively, as well as an eMMC flash module with a capacity of 32, 256 and 64 GB. Orin variants are equipped with 8 and 16 GB of LPDDR5 memory and a 256 GB NVMe SSD drive.
All servers contain ten computing nodes and an additional control node on the RK3588 chip. There is a slot for SSD with SATA-3 interface. The equipment includes two 10GbE (SFP+) and 1GbE (RJ45) network ports, a dedicated 1GbE network management port, an analog D-Sub connector (1080p60), two USB 3.0 ports, and a serial port. Air cooling is used. A 550 W unit is responsible for power supply. Operating temperature range – from 0 to +45 °C.
Image Source: Firefly
A cluster based on CSB1-N10S1688 provides AI performance of up to 160 TOPS on INT8 operations. In the case of CSB1-N10R3588 and CSB1-N10R3576, this figure is 60 TOPS. These three systems support working with large language models (LLM), as well as the TensorFlow, PyTorch, PaddlePaddle, ONNX, and Caffe frameworks. In the case of CSB1-N10NOrinNano and CSB1-N10NOrinNX, performance reaches 400 and 1000 TOPS. It talks about support for LLaMa3 and Phi-3 Mini, frameworks TensorFlow, PyTorch, Matlab, etc. Prices range from $2059 to $14,709.