Specifications
Product
Firefly CSB1-N10 series 1U AI cluster server
Architecture
10 compute nodes + 1 control node
SoC Options
SOPHON BM1688, Rockchip RK3588, Rockchip RK3576, NVIDIA Jetson Orin Nano, NVIDIA Jetson Orin NX
Control Node
Rockchip RK3588 @ 2.4 GHz
AI Compute — BM1688
160 TOPS INT8 (octa-core BM1688 @ 1.6 GHz)
AI Compute — RK3588
60 TOPS INT8 (octa-core @ 2.4 GHz)
AI Compute — RK3576
400 TOPS INT8 (octa-core @ 2.2 GHz)
AI Compute — Orin Nano
100 TOPS INT8 (hexa-core @ 1.5 GHz)
AI Compute — Orin NX
1000 TOPS INT8 (octa-core @ 2.0 GHz)
RAM per Node
BM1688 8 GB LPDDR4 (to 16 GB); RK3588 16 GB LPDDR4 (to 32 GB); RK3576 8 GB LPDDR4 (to 16 GB); Orin Nano 8 GB LPDDR5; Orin NX 16 GB LPDDR5
Storage per Node
32 GB eMMC (256 GB option) / 256 GB eMMC / 64 GB eMMC / 256 GB PCIe NVMe SSD
Video — BM1688/RK3576
160 channels H.265/H.264 1080p30 decode, 100 channels encode
Video — Orin
10 channels H.265 8K30 decode, 10 channels encode
Storage Expansion
SATA 3.0 / SSD drive slot
Ethernet
2× 10G Ethernet (SFP+), 2× Gigabit Ethernet (RJ45), 1× GbE RJ45 for management/BMC
Serial
1× RS-232 DB9 @ 115200, 1× RS-485 DB9 @ 115200
Buttons
Reset, UID, power
Power Supply
550 W (non-hot-swappable)
Form Factor
1U rackmount
Dimensions
420 × 421.3 × 44.4 mm
Target Workloads
NLP, robotics, image generation, private cloud inference
Overview
The Firefly CSB1-N10 series is a 1U rackmount AI cluster server built from ten compute nodes plus one control node. Rather than a single large accelerator, it aggregates ten low-power edge-class SoCs behind a 10GbE fabric, which produces a different performance and cost profile from a GPU server: 60 to 1000 TOPS INT8 depending on node choice, but spread across ten independent failure domains and ten independently addressable workloads.
Five node configurations are offered. The SOPHON BM1688 build gives 160 TOPS INT8 with strong video throughput — 160 channels of H.265/H.264 1080p30 decode — making it the video-analytics configuration. The Rockchip RK3588 and RK3576 options sit at 60 and 400 TOPS respectively as the general-purpose and mid-tier choices. The two NVIDIA builds use Jetson Orin Nano for 100 TOPS and Jetson Orin NX for 1000 TOPS, and both carry the Orin encode/decode block that handles ten channels of 8K30 H.265 in each direction.
The control node is an RK3588 at 2.4 GHz, and networking is two 10GbE SFP+ ports with two additional Gigabit RJ45 ports plus a dedicated management/BMC link. Each node has its own RAM and storage — from 8 GB LPDDR4 with 32 GB eMMC at the low end to 16 GB LPDDR5 with a 256 GB NVMe SSD on the Orin NX build — and the chassis adds a SATA 3.0 drive slot, RS-232 and RS-485 on DB9, and a 550 W non-hot-swappable supply in a 420 × 421.3 × 44.4 mm 1U frame.
Key Benefits
Ten independent inference nodes in 1U. Each node is its own failure domain and its own workload, rather than one accelerator shared across everything. 60 to 1000 TOPS on one platform. Five node options let one chassis design scale from a 60 TOPS video appliance to a 1000 TOPS inference cluster. Two 10GbE SFP+ ports. A real fabric for distributing work and results between nodes and the upstream network. High-density video analytics. The BM1688 configuration decodes 160 channels of 1080p30 H.265/H.264 on its own.
Applications
Private and on-premises AI inference clouds, video-analytics and smart-city deployments, natural-language processing services, image generation and diffusion workloads, robotics fleet inference, edge data-centre and micro-datacentre build-outs, research and university compute clusters.
Request a Quote — FIREFLY CSB1-N10 SERIES — 1U AI CLUSTER SERVERS FROM 60 TO 1000 TOPS
QS Compute — global B2B supply of AI computing hardware, edge AI systems and accelerators. Volume pricing, 15-day sample lead time.
Get Your Quote →