Specifications
AI Processor
SOPHON SG2300x TPU
INT8 Performance
24 TOPS
FP16 / BF16
12 TOPS
FP32
2 TOPS
Cascade
2x module cascade → up to 48 TOPS aggregate
Host Interface
PCIe x4, RC or EP mode
Video Codec
32 channels HD hardware decoding
Software
Radxa + SOPHON AI SDK (BMNNSDK) — driver, compiler and inference deployment toolchain; Radxa Model Zoo
OS
Ubuntu, CASA OS with web management panel
Development
Open source with hardware reference design available for secondary development
Target Workloads
Private GPT, Stable Diffusion, ChatDoc, public security, smart healthcare, industrial AI
Availability
Approved-partner network
Overview
The Radxa AICore SG2300x is a compact AI compute module built around the SOPHON SG2300x tensor processing unit. It delivers 24 TOPS at INT8, 12 TOPS at FP16/BF16 and 2 TOPS at FP32, and can be cascaded two-deep for up to 48 TOPS of aggregate compute. The module exposes PCIe x4 in either root-complex or endpoint mode, so it can act as an accelerator card in a host system or as the primary compute of a standalone edge device. It also includes 32-channel HD hardware video decoding, making it suitable for dense multi-camera analytics as well as generative workloads. Radxa ships the module with an open-source design and a hardware reference design for secondary development, plus the SOPHON BMNNSDK toolchain and the Radxa Model Zoo; supported operating systems are Ubuntu and CASA OS with a web management panel. The module is available through Radxa's approved partner network.
Key Benefits
Generative AI at module scale: the SG2300x is specified for private GPT, Stable Diffusion and ChatDoc deployments rather than classification-only pipelines, so it can host a local LLM or image-generation service without a GPU. Cascading for headroom: two modules in cascade deliver 48 TOPS, letting a design start small and scale inference within the same carrier. Flexible system role: PCIe x4 RC/EP support means the same module is either an accelerator inside an industrial PC or the compute core of a purpose-built edge appliance. Lower integration risk: open-source design, available hardware reference design, and the standard SOPHON BMNNSDK compiler and runtime cover model optimisation through to deployment. Camera density: 32 channels of HD hardware decode covers multi-camera video analytics on a single module.Applications
Private on-premise GPT and document Q&A appliances, local Stable Diffusion image generation, public security and video surveillance analytics, smart healthcare imaging, industrial AI inference, and multi-camera edge video processing.
Request a Quote — RADXA AICORE SG2300X — 24 TOPS SOPHON EDGE AI COMPUTE MODULE
QS Compute — global B2B supply of AI computing hardware, edge AI systems and accelerators. Volume pricing, 15-day sample lead time.
Get Your Quote →