Firefly's CSC2-N48SPK3 Lands — A 48-Node RISC-V AI Server With 2,880 TOPS in a Single 2U Chassis

Firefly Just Built the First Genuinely Practical Rack-Scale RISC-V AI Server

The CSC2-N48SPK3 landed on May 20, 2026 — and for anyone tracking the maturation of RISC-V from a curious hobby ISA into a serious data-center contender, this is the platform that crosses the threshold. Firefly's new 2U rack-mounted RISC-V AI server packs 48 separate compute nodes into a single chassis, each built around a SpacemiT K3 octa-core RISC-V SoC clocked at 2.4 GHz, with up to 32 GB of LPDDR5 memory and 128 GB of UFS storage per node. Aggregate AI performance hits 2,880 TOPS of INT4 compute. And every node gets a dedicated PCIe NVMe SSD slot — 48 of them, supporting up to 768 TB of total capacity when populated with 16 TB drives.

For homelab builders dreaming about a serious RISC-V cluster, system integrators thinking through native-RISC-V software workloads, and enterprise architects looking for an alternative to the dominant x86 and Arm server ISAs, the CSC2-N48SPK3 is the kind of platform you build around — not just look at. The benchmark for "is RISC-V production-ready in the data center" is the existence of platforms like this. It now exists. Starting price: $38,829.

What Makes the CSC2-N48SPK3 a Real RISC-V Milestone

The structural pitch is that the CSC2-N48SPK3 is the first dense, rack-mountable RISC-V AI server where every component of the design has been engineered around the assumption that RISC-V is the native ISA. The K3 nodes are RISC-V. The compute fabric is RISC-V. The storage architecture, the AI accelerator topology, and the network topology are all sized for RISC-V workload patterns. A central Rockchip RK3588 acts as the orchestration controller, but the heavy lifting happens on the 48 K3 nodes that define the server's character.

The K3 Octa-Core Node Is the Atomic Unit

Each compute node inside the CSC2-N48SPK3 is built around a SpacemiT K3 SoC: eight RISC-V cores at 2.4 GHz, RVV 1.0 vector extensions, 130,000 DMIPS of general-purpose compute, and a single-core SpecInt2006 result above 9.0 per GHz. Each node carries up to 32 GB of LPDDR5 memory and 128 GB of onboard UFS. The aggregate per chassis is 384 RISC-V CPU cores, up to 1.5 TB of LPDDR5 memory, and 6 TB of UFS — before you even count the 48 PCIe NVMe SSD slots that sit alongside the compute fabric.

The 60 TOPS Per Node × 48 = 2,880 TOPS Topology

The AI performance number — 2,880 TOPS of INT4 — comes from 60 TOPS per K3 node times 48 nodes. That distribution matters more than the headline figure. A topology where AI compute is co-located with each compute node makes inference workloads naturally embarrassingly parallel. The same multi-tenant inference scenario that requires careful workload partitioning on a single fat GPU server runs natively on the CSC2-N48SPK3 by just assigning each tenant a node.

Why Native RISC-V Compilation Is the Operational Differentiator

The single most concrete performance comparison Firefly published with the launch is the Linux kernel compilation benchmark: 22 minutes on the K3 RISC-V hardware versus nearly 3 hours on x86 Xeon emulation of RISC-V workloads. That is roughly 7× faster — and it captures the structural argument for native RISC-V infrastructure. The companies building RISC-V software stacks today have been compiling against emulators because production RISC-V hardware was rare. Replacing the emulation step with native silicon collapses the development loop dramatically, and the kernel compile result is the easiest way to see that on a single benchmark.

Software Optimization Loop Tightens Dramatically

For the broader RISC-V software ecosystem — Linux distributions, runtime libraries, ML frameworks, language runtimes — the CSC2-N48SPK3 is the kind of dense, available, on-premise platform that lets a small team iterate fast. Native compilation on real silicon at scale. Production-grade memory and storage. A standardized rack form factor. Those are the operational requirements for getting RISC-V software polished to the point where it competes with x86 and Arm in production. Firefly's box is the platform that meets those requirements at a price point a serious team can budget for.

How the CSC2-N48SPK3 Fits the Wider Edge AI and Homelab Landscape

The mini-computer and edge AI space has been on a steep maturation curve in 2026. Single-board RISC-V boards have been around for years, but the gap between a $100 SBC and a serious rack-mounted AI server has been the missing rung. The CSC2-N48SPK3 fills that rung. For homelab enthusiasts who have been clustering Raspberry Pi 5s or SpacemiT-based boards for fun, this is the platform that scales that experimentation up to a production-comparable footprint. For commercial customers, it is the dense RISC-V appliance that pairs with the recently-shipped Firefly AIBOX-K3 edge node to form a tier-1-to-tier-3 native RISC-V deployment stack.

48 NVMe Slots Is a Storage Story Worth Calling Out

A detail that often gets buried in compute-dense server specs but matters here: each of the 48 nodes gets a dedicated M.2 2280 PCIe NVMe slot. With 16 TB drives, that scales to 768 TB of total capacity in a single 2U chassis. For RISC-V workloads that need significant data locality — RAG inference, vector database serving, genomics, video processing — the storage density is part of the structural appeal. Compute and storage scale together inside the chassis, with each node owning its own local NVMe alongside its CPU and AI accelerator.

10GbE Connectivity Plus Management Networking

Networking is built for the workload patterns Firefly clearly has in mind. Four 10GbE SFP+ ports provide the high-bandwidth path to the rest of the rack. A separate gigabit management network handles orchestration traffic, IPMI-equivalent operations, and the central RK3588 controller's command-and-control fabric. That two-tier network topology is the standard data-center pattern for separating workload traffic from management traffic, and its presence on the CSC2-N48SPK3 is part of what makes the platform feel data-center-native rather than homelab-improvised.

What the $38,829 Price Tag Tells the Market

A platform like the CSC2-N48SPK3 is not a hobbyist board — at $38,829 starting, it is firmly enterprise-priced. But for a 48-node RISC-V server with 2,880 TOPS, that price point is competitive with comparable x86 or Arm options when total cost of ownership is calculated against native ISA compatibility for RISC-V software teams. The structural value is that Firefly has priced the platform to be deployable for a serious RISC-V engineering team or a forward-looking enterprise — not for a curious individual. That is the right place to land a category-defining product in 2026.

What to Watch Next

The CSC2-N48SPK3 is the kind of platform launch that opens follow-on questions. How quickly does mainline Linux, the major ML frameworks, and the standard cloud-native runtimes optimize for the SpacemiT K3 microarchitecture? Which enterprises are first to publicly deploy production RISC-V infrastructure? Does Firefly extend the CSC2 family with denser, higher-clock, or higher-TOPS configurations? And — the long-term question — does the existence of native RISC-V AI servers at this density pull more of the ML framework ecosystem onto native RISC-V kernel paths? Those answers will play out across the rest of 2026 and 2027. For now, the announcement marks the moment RISC-V crossed from interesting-hobby ISA to credible data-center contender.

Sources: CNX Software, May 20, 2026; Firefly product release, May 20, 2026; LinuxGizmos, May 2026; sbc.compare, May 2026.