The SC3 accelerator card is equipped with a BM1682 chip, which provides 3TFLOPS single-precision floating-point computing capability. Its actual computing power utilization rate is significantly higher than its competitors. SC3 provides two product forms of active/passive heat dissipation, which can be deployed on servers or industrial computers in the cloud and on the edge based on demand. It is suitable for business scenarios that have strict requirements on computing accuracy, such as industrial, medical, and dangerous goods management.
Each BM1682 chip has 64 NPU processing units, and each NPU has 32 EU arithmetic units. A single BM1682 chip can provide up to 3TFLOPs of single-precision peak computing power. At the same time, the chip has up to 16MB of on-chip SRAM, which can greatly reduce data handling during model calculation, improve performance and reduce power consumption.
High-performance deep learning accelerator card with completely independent intellectual property rights
On-chip hardware decoding engine supports HD video stream decoding from 1080P@240fps to 4K@60fps
Rich tool chain, support Caffe / TensorFlow / Pytorch / Mxnet and other deep learning frameworks
Passed CE / FCC and other international standard certification
Support PCIE 3.0 interface, compatible with mainstream x86 servers, easy to apply and expand
Passive cooling, fanless design
Sophon SC3 deep learning accelerator card can be used in various artificial intelligence, machine vision, high-performance computing environments, supporting facial feature detection, extraction, tracking, recognition, comparison, machine vision, and video structured analysis and processing Video structured applications such as image search and track tracking.
BMNNSDK (BITMAIN Neural Network SDK) one-stop toolkit provides a series of software tools including the underlying driver environment, compiler and inference deployment tool. The easy-to-use and convenient toolkit covers the model optimization, efficient runtime support and other capabilities required for neural network inference. It provides easy-to-use and efficient full-stack solutions for the development and deployment of deep learning applications. BMNNSDK minimizes the development cycle and cost of algorithms and software. Users can quickly deploy deep learning algorithms on various AI hardware products of SOPHGO to facilitate intelligent applications.
TPU architecture
Sophon
NPU Core number
64 Cores
Core Frequency
750Mhz
Single-Precision Performance(FP32)
3 TFLOPS
System Interface
PCI Express 3.0 x8
DDR Memory
8GB
On-Chip SRAM
16MB
TDP
65W
Thermal Solution
Passive(Fanless)
Video Decoder format
H.264/H.265 / HEVC / MPEG1 / 2 / 4 / DivX / XviD / H.263 / VC-1 / Sorenson / VP8 / AVS
Video decoder performance
1080p @ 240fps or 4K @60fps
DL Framework
Caffe/TensorFlow/Pytorch/Mxnet
OS Support
Ubuntu16.04 / CentOS7.4/Debian9.4
Operating Environment Temperature
0°C~55°C
Operating Environment Humidity
5%~95%RH
Storage Temperature
-45°C~75°C
Storage Humidity
5%~95%RH
Form Factor Length*Height*Thickness
217.21*125.44*21.59(mm)