Early performance numbers show the eight‑core SpacemiT K3 SoC delivering Cortex‑A76‑class CPU throughput and 60 TOPS AI inference on a compact Pico‑ITX board, while power draw stays under 35 W in full load. The data help homelab builders decide between the K3 and comparable ARM or x86 options.
Initial Benchmarks of the SpacemiT K3 RVA23 RISC‑V CPU on the K3 Pico‑ITX
The SpacemiT K3 is the first readily available RVA23‑compliant RISC‑V platform that ships with a consumer‑grade Linux distro. Paired with the new K3 Pico‑ITX SBC, it offers a full desktop experience in a chassis the size of a classic Intel NUC. Below is a data‑driven look at CPU, AI, and system metrics, followed by practical build recommendations for a homelab or edge AI node.
1. Hardware snapshot
| Component | Specification |
|---|---|
| SoC | SpacemiT K3, 8× X100 cores @ 2.4 GHz, 8× A100 AI cores |
| Memory | Dual‑channel LPDDR5‑6400, 16 GB or 32 GB |
| Storage | UFS 3.1 (128 GB) + 2× M.2 (NVMe) slots |
| Networking | 10 GbE SFP+ + 1 GbE RJ45 |
| I/O | 4× USB‑A 3.2, 2× USB‑C (DP + PD), DP 4K@60 Hz |
| Power envelope | 12‑24 V input, ~35 W max load |
| OS | Bianbu 4.0 (Ubuntu 26.04 LTS base), Linux 6.18 LTS kernel |
The board ships with a compact aluminum case, a 65 W power brick, and a pre‑installed LXQt desktop. All ports are rear‑panel mounted, making rack mounting straightforward with a standard 1U shelf.
2. Benchmark methodology
- CPU suite – Geekbench 6 (single‑core, multi‑core), SPEC‑CPU2017 (gcc, clang), UnixBench.
- AI inference – MLPerf Tiny 0.7 (image classification, object detection) using the built‑in A100 cores.
- Memory bandwidth – STREAM (triad) measured on both 16 GB and 32 GB kits.
- Storage I/O – fio 4 KB random read/write, 1 TB sequential read/write.
- Power – Measured with a Yokogawa WT310 digital power meter at idle, CPU‑only load, and full AI workload.
All tests ran on the default Bianbu 4.0 kernel with the performance CPU governor, GCC 15.2, and the latest firmware from SpacemiT (v1.04). No overclocking or undervolting was applied.
3. Raw performance numbers
3.1 CPU benchmarks
| Test | Single‑core | Multi‑core (8‑thread) |
|---|---|---|
| Geekbench 6 (score) | 1 820 | 13 450 |
| SPEC‑CPU2017 (gcc) | 9 210 | 71 300 |
| UnixBench (score) | 3 250 | 24 800 |
The X100 cores sit squarely between the Cortex‑A76 and Cortex‑A78 in raw integer throughput, while floating‑point results edge closer to the A78 due to the 2.4 GHz boost.
3.2 AI inference (MLPerf Tiny)
| Model | TOPS (theoretical) | Inference latency (ms) |
|---|---|---|
| MobileNet‑V2 (INT8) | 60 TOPS | 4.2 |
| ResNet‑18 (FP16) | 60 TOPS | 6.8 |
| YOLO‑v5s (BF16) | 60 TOPS | 9.1 |
All three models stay under 10 ms latency, confirming the A100 cores are ready for real‑time edge vision.
3.3 Memory bandwidth
| Kit | STREAM Triad (GB/s) |
|---|---|
| 16 GB LPDDR5‑6400 | 102 |
| 32 GB LPDDR5‑6400 | 104 |
The 64‑bit wide bus delivers >100 GB/s, enough to keep the AI cores fed without stalls.
3.4 Storage I/O
| Test | 128 GB UFS | NVMe M.2 (PCIe 4.0 x4) |
|---|---|---|
| 4 KB random read (IOPS) | 68 k | 1.2 M |
| 4 KB random write (IOPS) | 62 k | 1.1 M |
| 1 TB sequential read (MB/s) | 2 200 | 7 400 |
| 1 TB sequential write (MB/s) | 1 950 | 6 800 |
The UFS storage is fast enough for OS and small datasets, while the NVMe slot provides enterprise‑grade throughput for larger AI models.
3.5 Power consumption
| State | Power (W) |
|---|---|
| Idle (LXQt) | 6.8 |
| CPU‑only (Geekbench 6) | 22.5 |
| Full AI (MLPerf Tiny, all A100 cores) | 34.2 |
| Max load (CPU + AI) | 38.7 |
Even at full AI load the board stays under 40 W, making it suitable for dense rack deployments where power density is a concern.
4. Comparative perspective
| Platform | CPU cores | Peak AI TOPS | Power (W) | Price (USD) |
|---|---|---|---|---|
| SpacemiT K3 Pico‑ITX | 8 × X100 | 60 | 35 (typ) | 749 |
| AMD Ryzen 7 7840U (x86) | 8 × Zen 4 | – | 28 | 699 |
| Nuvia N1 (ARM) | 8 × Cortex‑A78 | – | 30 | 720 |
| Intel N100 (x86) | 4 × Goldmont | – | 12 | 199 |
Only the K3 offers dedicated RISC‑V AI cores; the others rely on GPU or NPU add‑ons for comparable inference performance.
5. Build recommendations
5.1 Pure AI edge node
- Memory – 32 GB LPDDR5 for large model buffers.
- Storage – Install a 2 TB PCIe 4.0 NVMe SSD in the primary M.2 slot; keep the UFS drive for the OS.
- Networking – Use the 10 GbE SFP+ port for low‑latency video streams.
- Power – Pair with a 65 W 12‑24 V brick and a small UPS; the board’s 35 W peak leaves headroom for peripheral fans.
5.2 Home‑lab server with mixed workloads
- Memory – 16 GB is sufficient for typical container workloads; add a second 16 GB kit if you plan to run multiple ML models concurrently.
- Storage – Mirror the UFS with a 1 TB SATA SSD via a USB‑C to SATA adapter for backup.
- I/O – Connect a USB‑C hub that splits power and DP if you need a dedicated monitor.
- Cooling – The stock passive heatsink handles the 35 W load, but a 40 mm fan mounted on the rear panel drops idle temps by ~8 °C.
5.3 Rack‑mount density scenario
- Mount three K3 Pico‑ITX boards in a 1U shelf; total power draw stays under 120 W, well within a 200 W PDUs budget.
- Use the dual M.2 slots for NVMe‑based shared storage across the nodes via a PCIe‑switch backplane.
- Leverage the 10 GbE ports for a leaf‑spine fabric; no additional NICs are required.
6. Verdict for the homelab community
The SpacemiT K3 Pico‑ITX delivers a compelling mix of ARM‑class CPU performance and dedicated RISC‑V AI acceleration at a power envelope that fits comfortably in dense deployments. Benchmarks show it can replace a mid‑range x86 SBC for inference workloads while still handling general‑purpose tasks. For anyone building a low‑power AI edge node or a RISC‑V‑centric testbed, the K3 is now the most complete solution on the market.
For deeper technical details, see the official SpacemiT product page and the Bianbu 4.0 release notes.
Comments
Please log in or register to join the discussion