USB3 Accelerator Module Datasheet
USB3 Accelerator Module Datasheet
USB3 Accelerator Module Datasheet
Features
Description
The Coral Accelerator Module is a multi-chip module (MCM) designed to perform high-speed inferencing for machine
learning (ML) models. It includes the Edge TPU ML accelerator with integrated power control, and it can be connected over
a PCIe Gen2 x1, USB2, or USB3 (3.1 Gen 1) interface.
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.
Confidential: This is a special version of the Accelerator Module datasheet, intended only for select customers who want
to integrate the module using the USB 3.0 interface, which requires special design considerations.
Ordering information
See https://coral.ai/products/accelerator-module/.
Table of contents
Features 1
Description 1
Ordering information 1
Table of contents 2
1 Block diagram 3
2 Electrical characteristics 4
2.1 Recommended operating conditions 4
2.2 Absolute maximum ratings 4
2.3 Logic threshold levels 4
2.4 Power consumption 5
2.5 Peak performance 5
2.6 IO supplies 6
4 Application details 9
4.1 Example circuit designs 9
4.1.1 PCIe 9
4.1.2 USB 2.0 10
4.1.3 USB 3.0 11
4.2 Trace length compensation 12
4.3 Power-on sequence 13
4.3.1 PCIe 13
4.3.2 USB 2.0 and USB 3.0 13
4.4 Power delivery network design 14
4.5 Thermal management 14
4.5.1 Thermal limits and resistance 14
4.5.2 Temperature warnings and frequency scaling (PCIe only) 15
4.5.3 Fixed operating frequency (USB only) 15
4.6 Software requirements 15
5 Package information 16
5.1 Package and pin dimensions 16
5.2 Land pattern 17
5.3 Soldering recommendations 18
5.4 Tape and reel information 19
5.5 Weight 21
5.6 Storage conditions 22
6 Document revisions 22
1 Block diagram
2 Electrical characteristics
1
Case temperature is defined as the surface temperature of the module. For details, see 4.5 Thermal management.
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.
1
The maximum operating temperature of the case 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.
2
PMIC digital I/O pins: PGOOD4, PMIC_EN
3
Edge TPU digital I/O pins: USB_SEL, RST_L, INTR, CLKREQ_L, SD_ALARM
Low-level max (VOL) High-level min (VOH) Low-level max (VIL) High-level min (VIH)
1
PMIC digital I/O pins: PGOOD4, PMIC_EN
The power consumed by the 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 4. 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 by the Edge TPU 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 from a single Edge TPU. 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 4.4 Power delivery network design.
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
Table 5. Maximum current consumed by the module (for power supply design)
Parameter Max
Peak performance when the Edge TPU is running at the maximum operating frequency:
2.6 IO supplies
Note: USB3.0 designs that use I2C require option B (dual supply) as the PMIC and TPU I2C lines are tied together.
The Coral Accelerator Module has two distinct IO supply voltages: one for the PMIC (AON) and one for the Edge TPU. The
Edge TPU IO supply is always 1.8 V, generated internally from the 3.3 V main input source (VIN). This means that you have
two options for the module power supplies, depending on the AON supply provided:
In either scenario, the Edge TPU outputs (INTR, SD_ALARM) are always using 1.8 V output logic levels. If connected to a
host with different logic levels, they must be level shifted. However, the PMIC_EN input thresholds are always compatible
with 1.8 V CMOS signaling, regardless of the AON supply voltage.
4 Application details
You can integrate the Accelerator Module into a system design using either the PCIe or USB2 interface with very few
supporting components. The following diagrams show typical application circuits with either PCIe or USB2 interfaces.
Note: All ground terminals should be connected, especially the center contacts for thermal dissipation.
4.1.1 PCIe
Caution: The USB 3.0 PHY is very particular and we can't guarantee consistent operation with your specific design.
Due to the fact that USB 3.0 PHY on the Accelerator Module was tuned for short trace lengths, we can only support USB
3.0 on devices where the SoC and the Edge TPU are on the same PCB, with direct traces between the two. Typically, we
won't know whether the Accelerator Module works with USB 3.0 until the board is built and tested, and we are unable to
tune the PHY for your specific design.
Table 7 describes the high-speed signals that require each pair to have the same total trace length. Due to space
constraints, not all tracings in the module match for each pair, as indicated in the table. You must incorporate any necessary
length compensation into your hardware.
Table 7. Pins that must have matching pair lengths, and their internal trace lengths.
Pair Pin name Trace length (mils) Time delay (ps)
4.3.1 PCIe
Note: The 10 millisecond delay between PMIC_EN and RST_L is a conservative estimate. For a more immediate
response, monitor PGOOD4 and raise RST_L high 2 milliseconds after PGOOD4 is asserted. Otherwise, the illustrated
delay for RST_L should be based on the later of either PMIC_EN or VIN rising.
Caution: If you do not properly design your power delivery network (PDN) to handle peak currents from the Edge TPU, it
can easily overwhelm your system and cause brownouts.
As described in section 2.4 Power consumption, the current drawn by the Edge TPU is highly variable and depends on the
model being executed, so you must design your power supply based on peak power. Although the average current drawn
by the Edge TPU might seem low (less than 500 mA), it can 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. However, these numbers are representative of only the models tested at Google, and your numbers
will vary based on your models.
To properly design a PDN for this module, you must consider the current envelopes generated when executing the ML
models you'll use in production. The current drawn by the Edge TPU is typically in the form of a few high current peaks, the
number of which depends on the model. The burst of high current peaks is usually followed by relatively long periods of
inactivity at idle currents. The peaks repeat at regular intervals, depending on the model architecture and number of
inferences per second.
The variation of current profile between models makes it very difficult to design a PDN that works for all applications.
Ultimately, you must optimize your own PDN based on the models you will use.
In particular, you must fine-tune the loop response in the DC/DC converter so it can absorb the load transients caused by
sudden and extreme changes in load current. Likewise, your PDN should maintain a low voltage ripple on the rail and avoid
internal overcurrent protection or inductor saturation events. It's important that you validate VIN’s PDN performance, such
as ripple noise and load step response performance when running your production models.
For more information about how to achieve and test these requirements, refer to the application information from the
vendor that provides your DC/DC converter.
Power dissipation in the Accelerator Module depends on the operating frequency and computational load. As the Edge
TPU heats up, performance may be affected, so it's important you design your system to manage thermal variations.
Note: The information in section 2.4 Power consumption includes some sustained power values (table 4) that can help
you estimate long-term thermal dissipation. But be sure you perform your own measurements, because total power
consumption varies based on the model you're running and other device characteristics.
The case temperature Tc and the Edge TPU’s junction temperature Tj should stay below the maximum operating specs:
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.
When designing a cooling solution for the module, be sure you consider the thermal behavior of the package when
attached to a heatsink. For simulation purposes, you can model the module using these absolute thermal resistance
properties:
These values represent the temperature difference between the Edge TPU’s junction and the top/bottom surfaces for a
given power flow across the interface, respectively. Figure 7 illustrates these temperature limit locations.
Figure 7. Module cross-section showing Edge TPU junction and module thermal properties
To estimate the effectiveness of your cooling solution—and to calculate a total thermal resistance—you should model θj-c
and θj-b thermal impedances in series with the thermal impedances of your interface material and heatsink design.
The Edge TPU includes an internal temperature sensor to help you make power management decisions. If you're using
PCIe, you can manually read the temperature, configure parameters that specify when the INTR and SD_ALARM pins assert
based on the current Edge TPU junction temperature, and specify trip-points for dynamic frequency scaling (DFS).
If you connect the Coral Module using the USB interface, then the temperature readings and DFS functionality is not
available. Instead, the operating frequency is fixed and you must measure the system temperature yourself.
You can choose to run the Edge TPU at either the "maximum" (500 MHz) or "reduced" (250 MHz) operating frequency when
you install the Edge TPU runtime on the host system.
The Accelerator Module must be operated by the Edge TPU runtime and Coral PCIe driver, which are 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 Package information
To avoid a short circuit due to solder contact with the side shielding, be sure the solder for pads along the module
perimeter do not extend to the module outline.
The tape is wound clockwise, with feeding holes to the right side as the tape is pulled toward the user.
The cover tape and base tape are not adhered within the "no components" area for 250mm (min).
The peeling of force is 1.1N (max) in the direction of peeling, as shown in figure 15.
This product is rated to MSL 3. To ensure proper storage conditions, tape and reel must be sealed within the provided
anti-humidity plastic bag. The bag contains a desiccant and humidity indicator.
5.5 Weight
Taping 144 g
Reel 275 g
● Please use this product within 6 months of receipt. If unused for more than 6 months, you must verify the product is
ready for soldering.
● While still in the anti-humidity packaging, the product should be stored at an ambient temperature from 5 to 35 °C
and humidity from 20 to 70% RH. (Packing materials may deform at temperatures over 40 °C).
● The product must not be stored in a corrosive environment gas (such as Cl2, NH3, SO2, NOx).
● Avoid any mechanical shock, such as dropping the packaging materials.
Baking conditions:
● 125+5/-0 °C, 24 hours, 1 time
● Bake individual modules on a heat-resistant tray because the materials (base tape, reel tape and cover tape) are
not heat-resistant.
6 Document revisions
1.4.1 (March 2021) Added USB 3.0 information: Added "USB3 interface" column in table 6 and added section 4.1.3.
1.3 (September 2020) Fixed the name and time delay for PCIE_RX_N (table 7).
1.2 (August 2020) Changed max Edge TPU junction temperature (Tj) to 115 °C (was 125 °C, which is actually used
for HTOL and other qualifications).
1.0 (July 2020) Updated electrical characteristics, power consumption, thermal management, packaging specs,
and miscellaneous edits.