Cambricon MLU270-X Series User manual

Cambricon Technologies, MLU270-X Series Intelligent Processing Card User Manual
MLU270-X Series Intelligent Processing
Card User Manual
V0.9

Cambricon Technologies, MLU270-X Series Intelligent Processing Card User Manual
Revision History
Document Name
MLU270-X Series Intelligent Processing Card User Manual
Revision
V0.9
Created Date
2020.01.10
Revision History
No.
Date
Version
Notes
0
2020.01.10
V0.9
First Release
1
2
3
4
5
Technical supports contact: [e-mail]
www.cambricon.com

Cambricon Technologies, MLU270-X Series Intelligent Processing Card User Manual
Content
1. Product Brief...................................................................................................................................4
2. Product specifications................................................................................................................... 5
2.1 Performance specifications.................................................................................................5
2.2 Software specifications........................................................................................................5
2.3 Operating Environment...................................................................................................... 7
2.4 Mechanical Specifications...................................................................................................7
2.5 Package Size and Weight................................................................................................... 8
2.6 Thermal Specifications........................................................................................................ 8
2.7 Power Supplies Specifications............................................................................................9
3. Cambricon NeuWare Development Environment................................................................... 12
4. Compliance.................................................................................................................................. 12

Cambricon Technologies, MLU270-X Series Intelligent Processing Card User Manual
1. Product Brief
Figure.1 MLU270-X Series Intelligent Processing Card appearance
Specially designed MLU270-X series dedicated to AI inferences and data center
accelerations with high EER (Energy Efficiency Rate)
The SIYUAN 270 ASIC has been designed based on several innovative technologies in
the architecture of processors owned by Cambricon. Along with it, many up-to-date
features have been integrated into a standard FHFL PCIe card, which can be inserted into a
modern AI PC/server so as to provide a huge extension of calculation power of AI inferences.
MLU270-X series has a moderate TDP of 150W and it can provide a calculation power as
high as 4 times of the previous generation. It can be widely used in applications such as
vision, voice, natural language processing, legacy machine learning, and many other AI
scenarios, and it can be used in an AI inferences platform to make it work with even higher
EER.
Brand-new Cambricon MLUv02 Architecture
The MLUv02 architecture is not just a simple update from previous generation, but a
brand new design based on NOC (Network on Chip), which will guarantee the parallel
efficiency of execution of 16 NPU cluster within the SIYUN 270 ASIC. The dataflow will be
compressed within the chip by dedicated hardware engine and this will increase the volume
and bandwidth effectively. The new architecture can fully support all the AI accuracies such
as INT16, INT8, INT4, FP32, FP16, and provide necessary calculation power to many kinds of
Neural Network. In a word, the new architecture can provide both good universality and best
performances to customers at the same time.
One more step on the performances of inferences
When using INT8 accuracy for AI inference calculations, the performance of non-sparse
networks is improved up to 4 times as much as that of the previous generation. MLU270-X

Cambricon Technologies, MLU270-X Series Intelligent Processing Card User Manual
series can provide a great EER as high as 40 times of a normal CPU. There are embedded
newly designed code/decode hardware for videos and pictures, so when the system have to
handle this kind of tasks, MLU270-X series will decrease the load of CPU and the occupation
for PCIe bandwidth, so as to increase the performances even more.
Calculating flexibility and programmability
SIYUAN 270 ASIC can provide supports to many kinds of Neural Networks. Cambricon
NeuWare software stack can easily deploy inference environment.BANG Lang.CHENG
environment can be directly customized for computing resources ，meeting all kinds of AI
customization requirements,professional but not specialized.
2. Product specifications
2.1 Performance specifications
Table 1 MLU270-X Series Intelligent Processing Card Hardware Specifications
Card Type
MLU270-X5/X5K/XV5K
Processor Architecture
Cambricon MLUv02
Core Clock
MLU270-X5:1 GHz,
MLU270-X5K/XV5K:1.25 GHz
Peak INT8
MLU270-X5:128 TOPS（Dense）,
MLU270-X5K/XV5K:160 TOPS（Dense）
Calculation accuracy
INT16, INT8, INT4, FP32, FP16
Video Decoding Support
Yes
Total Memory Size
32GB
Memory Bus Width
256-bit
Memory Bandwidth
102.4 GB/s
System Interfaces
PCI Express Gen3, x16, Supporting Lane Reversal
PCIe ID
PCIe Vendor ID 0xCABC
PCIe Device ID 0x0270
PCIe Sub-Vendor ID 0xCABC
PCIe Sub-System ID 0x0014/0x0015/0x0016
Outline
111.15mm*167.5mm, Double slots
TDP
150W
ECC Supporting
Yes
Heat Dissipation
Passive
2.2 Software specifications
Table 2 MLU270-X Series Intelligent Processing Card Software Specifications
PCIe Base address
PF （1，64bit）:

Cambricon Technologies, MLU270-X Series Intelligent Processing Card User Manual
BAR0: 256MB prefetchable
BAR2: 64MB prefetchable
BAR4: 64MB prefetchable
VF（4，64bit）：
BAR0: 256MB prefetchable
BAR2: 64MB prefetchable
BAR4: 64MB prefetchable
SMBus（8bit Address）
0x8E(write） 0x8F（read）
SMBUS Register is 32-bit wide, and below describes how to read a register（S:Slave，M:Master）：
Table 3 SMBUS Registers Reading
Direction
Bits
Content
MS
1
START
MS
8
SLAVE ADDRESS（Write）
SM
1
ACK
MS
8
REGISTER ADDRESS
SM
1
ACK
MS
1
RE START
MS
8
SLAVE ADDRESS（Read）
SM
1
ACK
SM
8
DATA[7:0]
MS
1
ACK
SM
8
DATA[15:8]
MS
1
ACK
SM
8
DATA[23:16]
MS
1
ACK
SM
8
DATA[31:24]
MS
1
NACK
MS
1
STOP
Table 4 SMBUS Registers Description
Registers
Address
ACCESS
Description
Total Card Power
0x01
RO
Card Power consumption, Float Data, Unit W
Card Temperature
0x02
RO
Card Temperature, Float Data, Unit ℃
Chip Temperature
0x03
RO
Chip Temperature, Float Data, Unit ℃
power brake
0x05
WO
The main frequency is reduced to 25% of current
frequency when write 0x04,and restore to
pre-frequency when write 0x01
PCIE Vendor ID
and Device ID
0xA0
RO
[15:0] Vendor ID：0xCABC
[31:16] Device ID：0x0270
PCIE Sub-Vendor
0xA1
RO
[15:0] Sub-Vendor ID：0xCABC

Cambricon Technologies, MLU270-X Series Intelligent Processing Card User Manual
2.3 Operating Environment
Table 5 MLU270-X Series Intelligent Processing Card Working Environment
Operating Temperature
0℃～45°C
Storage Temperature
-40℃～75℃
Operating Humidity
5%—95% Relative Humidity
Storage Humidity
5%—95% Relative Humidity
2.4 Mechanical Specifications
ID and Sub-System
ID
[31:16] Sub-System ID：for example 0x0014
Device Name
0xF0
RO
Display device type，e.g 0X14 means device type is
XV5K
Vendor Name
0xF1
RO
Display vendor name
Hardware vision
0xF2
RO
Display hardware vision
Firmware Revision
0xF3
RO
Display Firmware Revision，e.g 0x02140016 means
the main chip version is 0x02,device type is
0x14,main version is 0x00,subversion is 0x1,patch
number is 0x6
Manufacturing Time
0xF4
RO
Dispaly manufacturing time,e.g 0x1811 means
manufactured in 2018.11
Device ID
0xF5
RO
Display Device ID，e.g 0x20023 means device id is
20023

Cambricon Technologies, MLU270-X Series Intelligent Processing Card User Manual
Figure. 2 MLU270-X Series Intelligent Processing Card Dimensions
2.5 Package Size and Weight
Table 6 MLU270-X Series Package Size and Weight
Type
Weight
Size
Notes
Single Card
480g
267mm*111.15mm*37.10mm
NA
Gift package
with single
card
920g
385mm*230mm*97mm
Including gift
package,MLU270-X series
card,expand stent,8-pin
dongle cable,PE bag,desiccant
Gift Box
10.3kg
600mm*400mm*253mm
6 gift packages each box
Notes：The weight is a measured value with a tolerance of +-10%
2.6 Thermal Specifications
2.6.1 MLU270-X Series Power and Temperature specifications
Table 7 MLU270-X Series Thermal Design specifications
Item
Parameters
TDP(Thermal Design Power)
150W
Recommended Operating Tj
0-90℃

Cambricon Technologies, MLU270-X Series Intelligent Processing Card User Manual
Slowdown Tj
92℃
Slowdown Rate
50%
Shutdown Tj
95℃
2.6.2 MLU270-X Series Airflow Directions Requirement
MLU270-X Series use a blower to cool the card，the airflow direction is shown below：
Figure.3 MLU270-X Series Airflow direction
2.7 Power Supplies Specifications
MLU270-X Series Intelligent Processing Card provides a 8-Pin Power Connector as
below：

Cambricon Technologies, MLU270-X Series Intelligent Processing Card User Manual
Figure.4 8-Pin Power supply Connector
Table8 Supported Auxiliary Power Connections
Users can use the CPU 8-Pin Power Socket within a Server directly, or use two 8-pin PCIe
Power Sockets, along with a One-To-Two Adapter Cable.
Table9 Adapter Cable Connections
Board Connector
Cable used
CPU 8-Pin
1× CPU 8-pin cable
CPU to PCIe 8-Pin Dongle Cable
2× PCIe 8-pin cable
2× PCIe 6-pin cable
1× PCIe 8-pin cable and 1× PCIe 6-pin cable
8-Pin CPU Power
connector
PCIe 8-Pin Socket 1
PCIe 8-Pin Socket 2
Colour
12V
5
3
Yellow
12V
6
1,2
Yellow
12V
7
3
Yellow
12V
8
1,2
Yellow
GND
1
7,8
Black
GND
2
5,6
Black
GND
3
7,8
Black

Cambricon Technologies, MLU270-X Series Intelligent Processing Card User Manual
Figure.5 Adapter Cable
Note: The first pin (Pin No.1) of every connector/socket should have obviously distinguishable
mark at the side.
Table 10Input Voltage Requirement
Power Supply
Min
Normal
Max
PCIe edge connector（12V）
11.04V
12V
12.96V
8pin connector（12V）
11.04V
12V
12.96V
PCIe edge connector（3V3）
3.0V
3.3V
3.63V
Table 11 Current Requirement
Power Supply
Peak Current
Moving Average
8pin connector（12V）
20A
200us
17A
1ms
13A
5ms
Table12 Power Capping
Power Capping Threshold
150W
Power Capping Response time （typical）
50ms
Power Capping Response time （max）
100ms
Table 13 Power Break
PB# PCIe pin assignment
B30
Power Brake response time（typical）
150us
PB# input insertion low time （min）
250ms
Power brake hardware slowdown factor
4x
GND
4
5,6
Black

Cambricon Technologies, MLU270-X Series Intelligent Processing Card User Manual
3. Cambricon NeuWare Development Environment
NeuWare can fully support all kinds of mainstream programming framework, such as
TensorFlow, Caffe, PyTorch, MXNet and so on.With above mentioned Programming Frameworks,
users can easily and conveniently develop and deploy their Deep Learning Applications on
Cambricon MLU270-X Series Intelligent Processing Cards.At the same time, NeuWare provides
complete runtime system and driver software to speed up the system integration procedure.
NeuWare also provides a full set of software tools such as Application Development,
Function Debugging and Performance Optimization. The Application Development Tools include
Machine Learning Library, Runtime Library, Compiler, Model retraining tools and
Domain-Specific (for example Video Analysis) SDK; The Function Debugging Tools can fulfill all
the requirements from different levels of Programming Framework and Function Library; The
Performance Optimization Tools include tools for performances analysis and system
monitoring.
Figure.6 Cambricon NeuWare
For more information,please visit www.cambricon.com
Tel: [phone]
E-mail: [e-mail]
Address: 11/F, Block D, Zhizhen Tower, No.7 Zhichun Road, Haidian District, Beijing, China
4. Compliance
The MLU270-X Series is compliant with the regulations listed in this section. Compliance
marks, including the FCC ID numbers, can be found on the labels of each devices.
United States
Federal Communications Commission (FCC)

Cambricon Technologies, MLU270-X Series Intelligent Processing Card User Manual
This device complies with Part 15 of the FCC Rules.
Operation is subject to the following two conditions: (1) This device may not cause harmful
interference, and (2) this device must accept any interference received, including interference
that may cause undesired operation.
This equipment has been tested and found to comply with the limits for a Class B digital
device, pursuant to Part 15 of the FCC Rules. These limits are designed to provide reasonable
protection against harmful interference in a residential installation. This equipment generates,
uses and can radiate radio frequency energy and, if not installed in accordance with the
instructions, may cause harmful interference to radio communications. However, there is no
guarantee that interference will not occur in a particular installation.
If this equipment does cause interference to radio or television reception, which can be
determined by turning the equipment off and on, the user is encouraged to try to correct the
interference by one or more of the following measures:
Reorient or relocate the receiving antenna
Increase the separation between the equipment and receiver
Connect the equipment into an outlet on a circuit different from that to which the
receiver is connected
Consult the dealer or an experienced radio/TV technician for help
Caution: Any changes or modifications not expressly approved by the party responsible for
compliance could void the user's authority to operate this equipment.
Underwriters Laboratories (UL)
UL Listed Product Logo for MLU270-X Series Intelligent Processing Cards，model name
MLU270-X.