Firefly CSC2-N48SPK3 — 2880 TOPS RISC-V AI Server with 48 SpacemiT K3 Nodes and 48 NVMe SSDs

Specifications

Overview

The Firefly CSC2-N48SPK3 is a 2U rack-mounted server that brings RISC-V compute density to production racks. Instead of conventional x86 or Arm blades, it carries up to 48 SpacemiT K3 compute nodes, each an octa-core X100 RISC-V SoC rated at up to 60 TOPS of sparse AI performance with 16GB or 32GB of LPDDR5 and 128GB of UFS storage.

Aggregated across a full 48-node configuration that reaches up to 2880 TOPS of sparse AI throughput. The platform supports the RISC-V RVV vector extension, native FP8 inference and multimodal acceleration, and Firefly rates local throughput above 10 tokens/s on 30B-parameter models running entirely on the nodes.

Storage is the second defining attribute: 48 M.2 PCIe slots let every compute node carry its own dedicated NVMe SSD, which at 16TB per drive scales to 768TB of local storage inside a single 2U chassis. A Rockchip RK3588 octa-core Arm processor with 8GB of RAM acts as the central control node, orchestrating the fleet.

Security and virtualisation are handled at the silicon level with three-level M/S/U privilege modes, built-in Spectre and Meltdown mitigations, and full hardware virtualisation through RVH1.0, RV AIA and RV IOMMU extensions. The system ships with a pre-installed Linux distribution and supports OBS and Koji build systems. Firefly reports the setup is dramatically faster than emulation, with single-node performance up to 7x higher than x86 QEMU — one benchmark compiled the Linux 6.18 kernel in 22 minutes 28 seconds versus 2 hours 50 minutes on an Intel Xeon Gold 6548Y+ under emulation.

Pricing starts at $38,829 for a 48-node configuration with 16GB RAM and 128GB storage per node plus an 8GB RK3588 control node; a 32GB-per-node version is also available. QS Compute quotes configurations and volume orders on request.

Key Benefits

Extreme node density — 48 independent RISC-V AI nodes in 2U. 2880 TOPS sparse aggregate across a full configuration. 768TB of local NVMe through 48 dedicated M.2 slots, removing shared-storage bottlenecks for distributed inference. Native FP8 and RVV vector support for modern quantised inference workloads. Silicon-level security with privilege modes, Spectre/Meltdown mitigations and RVH1.0 virtualisation. Build-farm ready — OBS and Koji support plus a documented 7x advantage over QEMU emulation.

Applications

RISC-V software build farms and CI infrastructure, distributed LLM inference on local 30B-parameter models, sovereign and architecture-diverse AI compute, storage-heavy edge clusters, research and university compute, and multi-tenant virtualised AI testbeds.

Related Products