Firefly CSC2-N48SPK3 – A 2880 TOPS RISC-V AI Server with 48 SpacemiT K3 Nodes, 48 NVMe SSDs

Key Takeaways

• 48 SpacemiT K3 nodes deliver 2,880 TOPS AI performance
• Each node offers up to 32 GB LPDDR5 and 128 GB UFS storage
• Optional 48 NVMe slots enable up to 768 TB total capacity
• Control node uses Rockchip RK3588 for unified management
• Pricing starts at $38,829, targeting enterprise AI workloads

Pulse Analysis

RISC‑V is moving from hobbyist boards to enterprise‑grade infrastructure, and Firefly’s CSC2‑N48SPK3 exemplifies that shift. By integrating 48 independent K3 SoCs—each an octa‑core X100M processor with 60 TOPS sparse AI capability—the server delivers a combined 2,880 TOPS, rivaling many GPU‑based solutions while keeping power and cooling footprints low. The design also includes a Rockchip RK3588 control node, providing a familiar Arm‑based management plane and simplifying firmware updates through its BMC interface.

Beyond raw performance, the CSC2‑N48SPK3 offers flexible storage and memory configurations that address the data‑intensive nature of modern AI inference. Each compute node ships with up to 32 GB of LPDDR5 and 128 GB of UFS, while 48 M.2 PCIe slots allow operators to populate up to 768 TB of NVMe storage using 8‑TB or 16‑TB drives. The platform supports the RVV vector extension, native FP8 inference, and hardware virtualization extensions, enabling efficient execution of large language models and secure multi‑tenant workloads. Benchmarks claim up to seven‑fold faster compile times compared with x86 QEMU emulation, underscoring the efficiency gains of native RISC‑V execution.

From a business perspective, the server’s $38,829 entry price positions it as a premium but competitive option for enterprises seeking to diversify their AI hardware stack. Its dense form factor and open‑source ISA reduce vendor lock‑in risk and may lower total cost of ownership over time, especially as the RISC‑V ecosystem matures. Early adopters in edge data centers, research labs, and AI‑focused cloud providers could leverage the platform’s scalability and low‑latency inference to differentiate services, while the broader market watches to see if RISC‑V can challenge entrenched x86 and ARM server offerings.