AI computing accelerator card SC5H

Sophon SC5H is equipped with a BM1684 chip, adopts the standard PCIE card size design of half height and half length, and is equipped with a side suction fan. It can be well adapted to various complex working conditions and can be used for distributed AI computing analysis on the edge side scenes. Its computing power performance is more than 10 times that of Intel E5 CPU; at the same time, the power consumption is reduced by more than 70% compared with similar products. Its typical power consumption during overall operation is less than 21W.

Provide 2.2T@FP32, 17.6T INT8, 35.2T INT8 (Winograd ON) super computing power

High-performance power consumption ratio, for applications with high computing power requirements at the edge

Support multiple precision calculations such as FP32 and INT8

38-channel HD video hard decoding capability, applicable to high-speed high-frame rate industrial cameras

2-channel HD video hard-coding capability, supporting 4K level semi-real-time encoding output

Video and picture decoding resolution range up to above 8K, suitable for all kinds of ultra-high-definition network cameras

Adapt to various local workstation environments, and domestic CPU systems such as Feiteng, Shenwei, etc.

Can be mixed with SC5 + and other accelerator cards and GPU cards of other brands to build a heterogeneous computing platform

Wide application and rich scenes

SC5H is mainly used for distributed AI computing and analysis scenarios on the edge side, such as traffic, urban management, smart communities, industrial inspection and other scenarios that require front-AI computing power; it can also be mixed with other models such as SC5 + for the same computing platform.

Easy to use and convenient, full stack efficient

BMNNSDK (BITMAIN Neural Network SDK) one-stop toolkit provides a series of software tools such as the underlying driver environment, compiler, inference deployment tool and so on. Easy to use and convenient, covering the model optimization, efficient runtime support and other capabilities required for the neural network inference stage, providing easy-to-use and efficient full-stack solutions for deep learning application development and deployment. BMNNSDK minimizes the development cycle and cost of algorithms and software. Users can quickly deploy deep learning algorithms on various AI hardware products of Fortune Group to facilitate intelligent applications.

Support mainstream programming framework

More

Specifications

AI Developer Portfolio

AI computing accelerator card

AI computing accelerator card

TPU core architecture

SOPHON

SOPHON

SOPHON

SOPHON

NPU core number

64

-

64

192

AI computing power

FP32(FLOPS)

2.2T

-

2.2T

6.6T

INT8(OPS) Winograd OFF

17.6T

-

17.6T

52.8T

INT8(OPS) Winograd ON

35.2T

-

35.2T

105.6T

CPU

ARM 8-core A53 @ 2.3GHz

-

ARM 8-core A53 @ 2.3GHz

3x ARM 8-core A53 @ 2.3GHz

VPU

Video decoding capability

H.264:1080P @960fps
H.265:1080P @1000fps

-

H.264:1080P @960fps
H.265:1080P @1000fps

H.264:1080P@2880fps
H.265:1080P@3000fps

Video decoding resolution

CIF / D1 / 720P / 1080P / 4K(3840×2160) / 8K(8192×4096)

-


CIF / D1 / 720P / 1080P / 4K(3840×2160) / 8K(8192×4096)

CIF / D1 / 720P / 1080P / 4K(3840×2160) / 8K(8192×4096)

Video encoding capability

H.264:1080P @70fps
H.265:1080P @60fps

-

H.264:1080P @70fps
H.265:1080P @60fps

H.264:1080P @210fps
H.265:1080P @180fps

Video encoding resolution

CIF / D1 / 720P / 1080P / 4K(3840×2160)

-

CIF / D1 / 720P / 1080P / 4K(3840×2160)

CIF / D1 / 720P / 1080P / 4K(3840×2160)

Video transcoding capability (1080P to CIF)

Max. 18 channels

-

Max. 18 channels

Max. 54 channels

JPU

JPEG image decoding capability

800 images / second @ 1080p

-

800 images / second @ 1080p

2400 images / second @ 1080p

Maximum resolution (pixels)

32768×32768

-

32768×32768

32768×32768

System interface

Data link

EP PCIE X8
RC PCIE X8

PCIE X2

PCIE X16

PCIE X8

Operating mode

EP+RC

SOC extension

EP

EP

Physical / power interface

PCIE X16

12VDC Jack

PCIE X16

PCIE X16

RAM

Standard configuration

12GB

-

12GB

36GB

Maximum capacity

16GB

-

16GB

48GB

Power consumption

30W MAX

No load: 6W
With load: 30W

30W MAX

75W MAX

Heat dissipation mode

active

-

active

passive

Working status display

N/A

LED x3 (power / hard disk / status)

LED x1

LED x1

External I/O expansion *

SD-Card

1

-

-

RESET Button

1

-

-

RJ45

2 *1000Base-T

-

-

USB

4

-

-

SATA

1

-

-

4G/LTE

1

-

-

micro USB

1

-

-

working temperature

0℃-55℃

-10℃-55℃

0℃-55℃

Deep learning framework

Caffe / TensorFlow / Pytorch / Mxnet / Darknet / Paddle

Operating system support

Ubuntu / CentOS / Debian

compatibility

Compatible with mainstream x86 architecture and ARM architecture servers

Localization support

Support domestic CPU system such as Feiteng, Shenwei, Zhaoxin, etc.; support domestic Linux operating system such as Kylin, Deepin, etc.; support domestic AI framework Paddle Lite

Length x height x width (including bracket)

200x111.2x19.8mm

206x28.5x59.5mm

169.1x68.9x19mm

169.1x68.9x19.5mm

* All external I/O expansion interfaces in the AI developer portfolio must be used with SC5-IO