Banana Pi BPI-CM6: RISC-V x CM4 Compatible x 2.0 TOPS AI Industrial Compute Module
Banana Pi launched the BPI-CM6 public sales. Powered by SpacemiT K1 (8-core RISC-V), it's an industrial compute module balancing 2.0 TOPS AI and Raspberry Pi CM4 compatibility. A solid choice for starting edge inference with RISC-V.
Banana Pi has started public sales of the Banana Pi BPI-CM6. Product info is available on the official page and docs.
Powered by the SpacemiT K1 (8-core 64-bit RISC-V), it is an industrial compute module that achieves both AI 2.0 TOPS and Raspberry Pi CM4 compatible sizing.
It’s a compelling choice for projects needing “custom I/O design” or “wide temperature ranges” for edge inference, industrial control, or education. However, users should be prepared for potential hurdles with the RISC-V software landscape.
8GB/16GB LPDDR4 (LPDDR4X also mentioned, unconfirmed)
Storage
eMMC 8/16/32/128GB (Default 16GB)
Wired LAN
RTL8211F PHY integrated (RGMII), 2x GbE RJ45 on IO board
Wireless/BT
SDIO WiFi/BT module (specs unconfirmed)
Video Output
HDMI 1.4, MIPI DSI
Camera
MIPI CSI x3
USB
USB 3.0, USB 2.0
Expansion
PCIe 2.1 (5-lane), 2x M.2 M-KEY on IO board
I/O
UART x10, etc.
Size
40x55mm (IO board: 56x85mm)
Temp
-40 to 85°C
Power
12V PD recommended for IO board operation
OS
Bianbu Linux, OpenWrt, etc.
8-core RISC-V (SpacemiT K1) and AI 2.0 TOPS packed into a CM4-compatible 40x55mm industrial compute module.
The IO board features field-ready I/O like 2x M.2 M-KEY and 2x GbE. Check official info for exact memory specs (LPDDR4 vs 4X).
Features
SoC / CPU
The core of BPI-CM6 is the SpacemiT K1 (8-core 64-bit RISC-V). Highlighting AI 2.0 TOPS suggests it’s designed with inference in mind, even on RISC-V.
It also includes a GPU (IMG BXE-2-32), making it a decent base for edge devices requiring GUI or light GPU tasks.
Heat and silence depend on the chassis/cooling design. With no official TDP listed, evaluation with the baseboard is necessary.
NPU / AI Performance
AI performance is “2.0 TOPS”. No specific LLM benchmarks or framework guarantees are found in primary sources, but official docs mention a “deepseek demo”, showing intent to support AI.
Rather than a “local GenAI box”, it’s positioned as “edge compute for field inference”. Realistically, it’s for small model inference, lightweight image classification, or voice keyword detection.
Memory & Storage
SKUs for varying memory (8/16GB) and eMMC (8-128GB) are listed. eMMC capacity is fixed at purchase; user RAM upgrades are unlikely.
Expansion storage is practical: The official IO board has 2x M.2 M-KEY slots, and docs guide on NVMe SSD installation.
Useful for NAS, log storage, edge AI datasets, or local build environments where eMMC falls short.
The rear connectors reflect the “mount on CM4 compatible baseboard” philosophy. No external ports means custom I/O design freedom, ideal for gateways, NAS, or industrial control units.
Network: RTL8211F PHY on module, 2x GbE RJ45 on IO board. “SDIO WiFi/BT module” is listed but specific Wi-Fi standards are unconfirmed.
Video: HDMI 1.4 and MIPI DSI. Max resolution/simultaneous outputs unconfirmed.
Power/OS: 12V PD recommended for IO board. OS support includes Bianbu Linux and OpenWrt.
Summary
BPI-CM6 is a niche-hitting industrial compute module offering “RISC-V, CM4 compatibility, and AI”.
Suited for: Edge inference, industrial control, custom NAS/gateways, education.
Not suited for: Those expecting a “run-anything” Raspberry Pi experience (due to RISC-V software maturity).
Price starts from around $67, making it a challenging and interesting option. Verify specs against official docs before purchase.
Banana Pi has released product information for their AI-focused core module "Banana Pi BPI-SM9". Equipped with SOPHGO BM1688 and rated at up to 16 TOPS (INT8), the official documentation emphasizes video analytics applications such as decoding and analyzing multiple HD video streams. However, there are inconsistencies in official sources regarding power (5V/12V), dimensions, and GbE port count, so verify whether you are purchasing the module alone, IO board, or kit before ordering.
In July 2025, Canonical officially announced that ESWIN Computing's EBC77 Series now supports Ubuntu 24.04 LTS. This SBC features a 64-bit OoO RISC-V CPU and an in-house developed NPU, making the RISC-V + Ubuntu + AI combination a notable highlight. However, detailed specifications for I/O, power, and other aspects remain unconfirmed in official primary sources, so additional research is needed before deployment.
PINE64 has listed the "StarPro64" on their official store. This RISC-V SBC is built around the ESWIN EIC7700X (Quad P550) and features an "AI-oriented" design with a stated NPU performance of 19.95 TOPS (INT8). However, the official Docs explicitly note that "the release is alpha and intended for development and testing," so evaluations assuming production-ready stability should be avoided at this stage.
Radxa has published the product page and official documentation for the "Radxa Cubie A7Z." This ultra-compact SBC packs an Allwinner A733 (Cortex-A76×2 + Cortex-A55×6) and a 3 TOPS-class NPU into a 65×30mm form factor, complete with Wi-Fi 6 and Micro HDMI. However, wired LAN and M.2 require expansion, and PCIe is limited to Gen3 x1 via FPC, so checking whether it fits your use case beforehand is recommended.
