SOPHON SG2300X SoC

Born for Generative AI

SOPHON SG2300X SoC

Features
Product Details
Block Diagram
Specification
Documentations

SG2300X - Born for Generative AI

Octa-core A53,
up to 2.3GHz
Support 12 channels HD
hardware encoding
Support 32 channels HD
hardware decoding
Support INT8,
FP16/BF16 and FP32
Mainstream frameworks Support :
Built-in 24TOPS@INT8 high-performance TPU (Tensor Processing Unit)

Features Overview

Powerful LLM Performance

Smooth AI Conversation Experience

Text to Image - Only 1.1s*

Stable Diffusion: 1.1s per image (512*512px)
Stable DiffusionXL: 7.4s per image (1024*1024px)

*This data is the result of utilizing LCM acceleration, with 4 steps, equivalent to the
effect of 20 steps in original SD, generating a 512*512 image.

Superior Multimedia Capabilities

12 Channel
Video Decoding

32 Channel
Video Decoding

1080P @750 IPS decode
1080P @250 IPS encode
JPEG
Resolution up to 32768*32768

1080P @800 IPS decode
1080P @300 IPS encode
H.264 & H.265
Resolution up to 7680*4320

TPU-MLIR

General Architecture TPU Compilation Toolchain

TPU-MLIR is a TPU compiler open-source project. The project provides a complete toolchain, which is used to transform the neural networks pre-trained under different frameworks into binary files bmodel that can be efficiently operated on TPU.

Software Optimized for Generative AI Applications

Powerful and Complete Tool Chain

Support for various deep learning framework models
Support for F32/F16/BF16/INT8
Built-in implementation of 150+ operators as well as user-defined operators
Support for merging preprocessing and postprocessing
Support for PTQ quantization and automatic mixed precision quantization tuning
Support for compile-time instantaneous data validation
Support for multi-core distributed inference
Support for large models: ChatGLM3-6B/LLaMA2/Qwen/Stable Diffusion, and so on
Open source and open-minded

Mature, Stable, Easy-to-use SDK with Excellent Compatibility

Providing a powerful development library (NN Compiler, BMCV, FFMPEG decoding interface, OpenCV encoding/decoding interface, OpenCV extensions)

Support for Cloud-edge-end Collaboration

Support for Docker containerization, Kubernetes extension management
Support for cloud-based management, updates, and upgrades of computing power and algorithms