M.

2 Accelerator datasheet
Version 1.5

Features

● Google Edge TPU ML accelerator

○ 4 TOPS total peak performance (int8)
○ 2 TOPS per watt
● Integrated power management
● PCIe Gen2 x1 interface
● Available in two M.2 form factors:
○ M.2-2230-A-E-S3 (A+E key)
○ M.2-2280-B-M-S3 (B+M key)
● Operating temp: -20 to +85 °C

Description

The Coral M.2 Accelerator is an M.2 module (either A+E or B+M key) that brings the Edge TPU ML accelerator to existing
systems and products.

The Edge TPU is a small ASIC designed by Google that accelerates TensorFlow Lite models in a power efficient manner: it's
capable of performing 4 trillion operations per second (4 TOPS), using 2 watts of power—that's 2 TOPS per watt. For
example, one Edge TPU can execute state-of-the-art mobile vision models such as MobileNet v2 at almost 400 frames per
second. This on-device ML processing reduces latency, increases data privacy, and removes the need for a constant
internet connection.

The M.2 form-factor allows you to add local ML acceleration to products such as embedded platforms, mini-PCs, and
industrial gateways that have a compatible M.2 card slot.

Ordering information

Part number Description

G650-04527-01 Coral M.2 Accelerator with A+E key

G650-04686-01 Coral M.2 Accelerator with B+M key

See https://coral.ai/products/m2-accelerator-ae.

Coral.ai | Copyright 2019 Google LLC.

M.2 Accelerator datasheet v1.5
G650-04527-01 / G650-04686-01

Table of contents

Features 1

Description 1

Ordering information 1

Table of contents 2

1 Specifications 3

2 Dimensions 4
2.1 A+E key dimensions 4
2.2 B+M key dimensions 4

3 Electrical characteristics 5
3.1 Absolute maximum ratings 5
3.2 Power consumption 5
3.3 Peak performance 6

4 Connector pinout 7
4.1 A+E key pinout 7
4.2 B+M key pinout 8

5 Application details 9
5.1 Software requirements 9
5.2 Power delivery and management 9
5.3 Thermal management 9
5.3.1 Thermal limits 10
5.3.2 Top-side cooling options 10
5.3.3 Bottom-side cooling options 11
5.3.4 Temperature warnings and frequency scaling 11

6 Document revisions 12

Coral.ai | Copyright 2019 Google LLC. 2

M.2 Accelerator datasheet v1.5
G650-04527-01 / G650-04686-01

1 Specifications

For in-depth mechanical details, refer to the PCI-SIG's PCI Express M.2 specification.

Table 1. Technical specifications

Physical specifications

Dimensions A+E key: 22.00 x 30.00 x 2.35 mm

B+M key: 22.00 x 80.00 x 2.35 mm

Weight A+E key: 3.1 g

B+M key: 5.8 g

Host interface

Hardware interface M.2 A+E key (M.2-2230-A-E-S3)

or M.2 B+M key (M.2-2280-B-M-S3)

Serial interface PCIe Gen2 x1

Operating voltage

DC supply 3.3 V +/- 10 %

Environmental

Storage temperature -40 to +85 °C

Operating temperature -20 to +85 °C 1

Relative humidity 0 to 90% (non-condensing)

Mechanical (non-op)

Shock 100 G, 11 ms (persistent)

1000 G, 0.5 ms (stress)
1000 G, 1.0 ms (stress)

Vibration 0.5 Grms, 5 - 500 Hz (persistent)

(random/sinusoidal) 3 Grms, 5 - 800 Hz (stress)

Compliance

Countries 2 Unit shipped as a component. Final system certification/compliance to be done by the

customer.

ESD 3 1 kV HBM, 250 V CDM

1
The max operating temperature depends on the power consumption and thermal management in your system.
2
We can provide a certification example to show that a reasonably designed system can meet certification requirements.
3
Always handle in a static safe environment.

Coral.ai | Copyright 2019 Google LLC. 3

M.2 Accelerator datasheet v1.5
G650-04527-01 / G650-04686-01

2 Dimensions

2.1 A+E key dimensions

● PCB width: 22.00 mm ± 0.15 mm

● PCB height: 30.00 mm ± 0.15 mm
● PCB thickness: 0.80 mm ± 0.08 mm
● Top-side component height: 1.55 mm ± 0.10 mm
● Bottom-side component height: 0 mm

For in-depth mechanical specs, refer to the PCI Express M.2 Specification.

Figure 1. A+E key card dimensions (in millimeters)

2.2 B+M key dimensions

● PCB width: 22.00 mm ± 0.15 mm

● PCB height: 80.00 mm ± 0.15 mm
● PCB thickness: 0.80 mm ± 0.05 mm
● Top-side component height: 1.55 mm ± 0.10 mm
● Bottom-side component height: 0 mm

For in-depth mechanical specs, refer to the PCI Express M.2 Specification.

Figure 2. B+M key card dimensions (in millimeters)

Coral.ai | Copyright 2019 Google LLC. 4

M.2 Accelerator datasheet v1.5
G650-04527-01 / G650-04686-01

3 Electrical characteristics

3.1 Absolute maximum ratings

Exceeding the absolute ratings can cease operation and possibly cause permanent damage. Exposure to absolute ratings
for extended periods of time can also adversely affect reliability.

Table 2. Absolute maximum ratings

Parameter Min Max

Storage temperature -40 °C 85 °C

Operating temperature -20 °C 85 °C 1

Edge TPU junction temperature (Tj) -40 °C 115 °C

Power supply (3.3 V) -0.3 V 6.0 V

1
The maximum operating temperature is for the entire assembly and assumes that the Edge TPU junction temperature (Tj)
does not exceed its absolute maximum rating, which depends on the power consumption and thermal management in your
system.

3.2 Power consumption

The power consumed by the card module depends on the ML model, the number of inferences per second, and the
operating frequency of the Edge TPU. For some examples of average sustained power consumption, see table 3. However,
it's also important that you consider the peak current transients that occur during inferencing.

The maximum current drawn by the Edge TPU is typically much higher than the average current. That's because when the
Edge TPU executes an ML model, it repeatedly activates a large number of arithmetic logic units (ALUs) simultaneously,
resulting in a pattern of brief but large current transients. Each model architecture also activates a different set and different
number of ALUs, meaning the magnitude and the shape of the transient current very much depends on the model.

Although the average current drawn from the 3.3V supply is typically less than 500 mA, brief current transients that occur
during inferencing can reach roughly 3 A. These spikes also occur suddenly: even a simple model can generate current
transients in excess of 1 A/μs. However, these numbers are representative of only the models tested at Google, and your
numbers will vary. To determine the actual peak supply current, you should observe the current when running the models
you will deploy in production.

For more information, see section 5.2 Power delivery and management.

Table 3. Examples of long-term sustained power during inferencing

Model 1 Low operating frequency Reduced operating frequency Max operating frequency
125 MHz 250 MHz 500 MHz

MobileNet v2 0.6 W (7.1 ms @ 141 fps) 0.9 W (3.9 ms @ 256 fps) 1.4 W (2.4 ms @ 416 fps)

Inception v3 0.5 W (58.7 ms @ 17 fps) 0.6 W (51.7 ms @ 19.3 fps) 0.7 W (48.2 ms @ 20.7 fps)
1
Pre-compiled models were tested using models_benchmark.cc

Typical idle power consumption is 375 - 400 mW.

Coral.ai | Copyright 2019 Google LLC. 5

M.2 Accelerator datasheet v1.5
G650-04527-01 / G650-04686-01

3.3 Peak performance

Peak performance when the Edge TPU is running at the maximum operating frequency:

● 4 trillion operations per second (TOPS), 8-bit fixed-point math

● 2 TOPS per watt

Coral.ai | Copyright 2019 Google LLC. 6

M.2 Accelerator datasheet v1.5
G650-04527-01 / G650-04686-01

4 Connector pinout

4.1 A+E key pinout

Table 4. A+E key card pinout

Bottom side pins Top side pins
Pin Signal Signal Pin
74 3.3V GND 75
72 3.3V NC 73
70 NC NC 71
68 NC GND 69
66 NC NC 67
64 NC NC 65
62 NC GND 63
60 NC NC 61
58 NC NC 59
56 NC GND 57
54 NC NC 55
52 PERST0# (3.3V) CLKREQ0# (3.3V) 53
50 NC GND 51
48 NC REFCLKn0 49
46 NC REFCLKp0 47
44 NC GND 45
42 NC PETn0 43
40 NC PETp0 41
38 NC GND 39
36 NC PERn0 37
34 NC PERp0 35
32 NC GND 33
30 Key E Slot Key E Slot 31
28 Key E Slot Key E Slot 29
26 Key E Slot Key E Slot 27
24 Key E Slot Key E Slot 25
22 NC NC 23
20 NC NC 21
18 GND NC 19
16 NC NC 17
14 Key A Slot Key A Slot 15
12 Key A Slot Key A Slot 13
10 Key A Slot Key A Slot 11
8 Key A Slot Key A Slot 9
6 NC GND 7
4 3.3V NC 5
2 3.3V NC 3
GND 1

Coral.ai | Copyright 2019 Google LLC. 7

M.2 Accelerator datasheet v1.5
G650-04527-01 / G650-04686-01

4.2 B+M key pinout

Table 5. B+M key card pinout

Bottom side pins Top side pins
Pin Signal Signal Pin
74 3.3V GND 75
72 3.3V GND 73
70 3.3V GND 71
68 NC NC 69
66 Key M Slot NC 67
64 Key M Slot Key M Slot 65
62 Key M Slot Key M Slot 63
60 Key M Slot Key M Slot 61
58 NC Key M Slot 59
56 NC GND 57
54 NC REFCLKp0 55
52 CLKREQ0# (3.3V) REFCLKn0 53
50 PERST0# (3.3V) GND 51
48 NC PERp0 49
46 NC PERn0 47
44 NC GND 45
42 NC PETp0 43
40 NC PETn0 41
38 NC GND 39
36 NC NC 37
34 NC NC 35
32 NC GND 33
30 NC NC 31
28 NC NC 29
26 NC GND 27
24 NC NC 25
22 NC NC 23
20 NC GND 21
18 Key B Slot Key B Slot 19
16 Key B Slot Key B Slot 17
14 Key B Slot Key B Slot 15
12 Key B Slot Key B Slot 13
10 NC NC 11
8 NC NC 9
6 NC NC 7
4 3.3V NC 5
2 3.3V GND 3
GND 1

Coral.ai | Copyright 2019 Google LLC. 8

M.2 Accelerator datasheet v1.5
G650-04527-01 / G650-04686-01

5 Application details

5.1 Software requirements

The M.2 Accelerator must be operated by the Edge TPU runtime and Coral PCIe driver, which is compatible with the
following systems:

● Linux:
○ 64-bit version of Debian 10 or Ubuntu 16.04 (or newer)
○ x86-64 or ARMv8 system architecture
● Windows:
○ 64-bit version of Windows 10
○ x86-64 system architecture
● All systems require support for MSI-X as defined in the PCI 3.0 specification

5.2 Power delivery and management

Caution: If you do not carefully consider the power demands of the ML models running the Edge TPU, along with the
ability of your host to handle the corresponding current transients, the peak currents might cause brownouts or other
abnormal behavior in the upstream power regulator.

As described in section 3.2 Power consumption, the current drawn by the Edge TPU is highly variable and depends on the
model being executed. Although the average current drawn by the Edge TPU might seem low (less than 500 mA), it can
repeatedly and rapidly spike up to 3 A, depending on the model you're running. These spikes also occur suddenly: even a
simple model can generate current transients in excess of 1 A/μs, which can last several tens of microseconds.

Ideally, your host system and M.2 socket can be designed to tolerate these higher currents, and your power supply can
provide fast transient response performance. Alternatively, you may use some software strategies to mitigate the effects of
the peak currents, such as underclocking the Edge TPU.

5.3 Thermal management

The Edge TPU dissipates power roughly proportional to its computational load. The resulting heat in the Edge TPU die must
be safely and reliably conducted away to avoid excessive die temperatures that can affect performance and reliability.

The primary heat-generating components on the card are the Edge TPU and power IC, located under the shield can as
indicated in figure 3. The shield can provides thermal coupling to these components with thermal pads—there is no air gap
between the components and the shield can. (For thermal resistance detail, see section 5.3.2 Top-side cooling options.)

During typical operation, approximately 90% of the system power dissipates from the Edge TPU, and the remaining 10%
dissipates from the power IC. Total power dissipation depends on the operating frequency and computational load.

Coral.ai | Copyright 2019 Google LLC. 9

M.2 Accelerator datasheet v1.5
G650-04527-01 / G650-04686-01

Figure 3. Approximate location of the power IC (PMIC) and Edge TPU (coupled with the shield can using thermal pads)

5.3.1 Thermal limits

The Edge TPU’s junction temperature Tj must stay below the maximum operating specification:

● Maximum Edge TPU junction temperature Tj: 115 °C

Warning: Exceeding the maximum temperature can result in permanent damage to the Edge TPU and surrounding
components, and can possibly cause fire and serious damage, injury, or death.

For information about how to read the Edge TPU temperature, see Manage the PCIe module temperature.

5.3.2 Top-side cooling options

To ensure successful long-term operation, you might want to add a cooling solution on the top-side of the card module, on
top of the shield can. When selecting a thermal solution for the top, consider the following thermal resistance properties
with the shield can in place:

● Edge TPU junction-to-shield-can thermal resistance θj-s: 11 °C/W

Although many applications can sustain proper thermal levels with the shield can in place, you can achieve higher thermal
dissipation (if necessary) by removing the shield can and placing a thermal solution in direct contact with the Edge TPU. If
you choose to do so, then consider the junction-to-case thermal resistance and component dimensions indicated in table 6.

Table 6. Thermal properties and dimensions for cooling solutions with the shield can removed
Component Top-face dimensions Top-face height from PCB Junction-to-case thermal resistance θj-c
(X-Y) (Z)

Edge TPU 5.0 x 5.0 mm 0.55 ± 0.03 mm 2.2 °C/W

Power IC 2.6 x 3.0 mm 0.48 ± 0.03 mm 0.5 °C/W

Shield can frame N/A ~1.35 mm N/A

Other N/A 1.00 ± 0.10 mm N/A

Notice that other top-side components are taller than the primary heat-producing components, so your heat sink or other
enclosure must clear those components. For improved thermal conductivity, consider adding metal stubs that extend from
the heat sink to the surface of the Edge TPU, and fill the remaining gap to the Edge TPU with a thermal coupling material.

Coral.ai | Copyright 2019 Google LLC. 10

M.2 Accelerator datasheet v1.5
G650-04527-01 / G650-04686-01

Caution: If you remove the shield can, it's important that your added heat sink or enclosure has sufficient clearance
above the tallest top-side components to prevent the risk of contact and electrical shorting.

If you remove the shield can, be sure to consider the distance between the PCB and heat sink or enclosure. This distance
determines the minimum allowable thermal pad thickness, as well as the maximum compressive force that can be exerted
on the card. To ensure safe operation, the sustained compressive pressure onto each component from the thermal pads
should not exceed 30 PSI (assuming there is an air gap below the card, and thermal pads on the entire top face of the Edge
TPU and power IC).

5.3.3 Bottom-side cooling options

A secondary thermal path for cooling the Edge TPU is a thermal epoxy or soft thermal pad on the underside of the card,
directly below the Edge TPU. This may dissipate some of the power through the card module and into the base PCB below.

The bottom-side cooling solution is less effective than the top-side solution and should be considered a supplemental
thermal path. In order to approximate the effectiveness of a bottom-side thermal path, you should use the junction-to-board
thermal resistance θj-b indicated in table 7.

Table 7. Thermal properties for bottom-side cooling solutions

Component Top-face dimensions (X-Y) Junction-to-board thermal resistance θj-b

Edge TPU 5.0 x 5.0 mm 15 °C/W 1

In this case, θj-b is the temperature difference between the Edge TPU junction and the surface of the card module when
1

measured from the bottom of the card, directly underneath the Edge TPU.

5.3.4 Temperature warnings and frequency scaling

The Edge TPU includes an internal temperature sensor to help you make power management decisions. You can manually
read the temperature, configure parameters that specify when the Edge TPU should shut down, and specify trip-points for
dynamic frequency scaling (DFS).

For details, read Manage the PCIe module temperature.

Coral.ai | Copyright 2019 Google LLC. 11

M.2 Accelerator datasheet v1.5
G650-04527-01 / G650-04686-01

6 Document revisions

Table 8. History of changes to this document

Version Changes

1.5 (August 2020) Changed max Edge TPU junction temperature (Tj) to 115 °C (was 125 °C, which is actually
used for HTOL and other qualifications).
Changed minimum operating temperature to -20 °C (was -40 °C).

1.4 (August 2020) Added information about power consumption and thermal management.
Updated operating temperature and system requirements.
Removed description of DFS; added a link to a more detailed app note.
Miscellaneous edits. Restructured document to match similar products.

1.3 (April 2020) Updated system architecture requirements

1.2 (December 2019) Revised dimensions and added tolerances

1.1 (October 2019) Added max power consumption

1.0 (August 2019) Initial release

Coral.ai | Copyright 2019 Google LLC. 12