Who is better: Google Coral USB accelerator or Intel NCS 2?
As artificial intelligence (AI) and machine learning (ML) gradually move from science fiction to real life, we now need a fast and convenient way to prototype this type of system. Although desktop computers can suffice to meet AI/ML operation requirements, even single-board computers like Raspberry Pi. But what if you only want a simple plug-in device to make your system run faster and more powerful?
Don’t worry, you actually have a variety of options, including the hardware USB Accelerator of Google’s Coral Edge TPU series (Coral USB Accelerator, hereinafter referred to as CUA) and Intel’s Neural Compute Stick 2 (neural computing stick NCS2). Both devices are computing devices that plug into the host computer via USB. NCS2 uses a visual processing unit (VPU), while Coral USB Accelerator uses a tensor processing unit (TPU), both of which are dedicated processing devices for machine learning. Today I will give you a comparison and evaluation: What is the difference between the two? As a developer, do you choose Coral or NCS2? Let’s see below.
-ML accelerator: Edge TPU ASIC (Application Specific Integrated Circuit) chip designed by Google. Provide high-performance ML inference (MobileNet V2 400 + fps, from the latest official update data) for TensorFlow Lite models.
-Support USB 3.1 port and cable (SuperSpeed, 5GB/s transfer speed)
-Dimensions: 30 x 65 x 8 mm
-Official price: $74.99
Intel Neural Compute Stick 2
-Processor: Intel Movidius Myriad X visual processing unit (VPU)
-USB 3.0 Type-A
-Dimensions: 72.5 x 27 x 14mm
-Official price: $87.99
1. Comparison of processor and acceleration performance
Unlike the way to compare traditional computer CPUs, comparing the details of each processor/accelerator is more subtle, depending on how you plan to use them. Although the output format is slightly different (each inference time and frames per second), we can still compare the two devices for some overall performance mode.
When evaluating AI models and hardware platforms for real-time deployment, the first thing to look at is how fast they are. In computer vision tasks, benchmarks are usually measured in frames per second (FPS). Higher numbers indicate better performance. For real-time video streaming, at least about 10 fps is required to make the video appear smooth.
Running performance: First, when CUA is added to the desktop CPU, the performance can be improved by about 10 times, and the running performance is relatively good. (Depending on the CPU model selected, the 10-fold performance slightly fluctuates up and down.) NCS2 cooperates with the older Atom processor to increase processing speed by nearly 7 times. However, when used with more powerful processors, the results presented by NCS2 are not surprising.
NCS2 can theoretically perform inference at 4 TOPS. Curiously, CUA also has the exact same rate, although both use different operations to perform ML. In addition, Intel claims that the performance of NCS2 is 8 times that of the first-generation neural computing stick. (If you want, you can choose NCS2 instead of the original neural computing stick, although the latter is much cheaper.)
NCS2 can use MobileNet-v2 to run a 30 FPS classification model, which is not bad. However, it is a bit difficult to detect objects at 11 FPS. A frame rate of about 10 FPS may not be sufficient for real-time object tracking, especially for high-speed motion. Many objects may be lost and developers need very good tracking algorithms to make up for this vulnerability. (Of course, the official benchmark results are not completely reliable. Usually, these companies compare their manually optimized software with competitors’ out-of-the-box models.)
User evaluation
Power consumption: NCS2 has lower power consumption. As far as CUA is concerned, the official does list the 0.5 watts required per TOPS. The user can also set the CUA to the default speed or maximum (2 times the default value) as needed.
It’s worth noting that Google’s official documentation does clearly remind you: When the device is running at maximum speed, the power transmission and maximum ambient temperature may burn your skin. So personally, unless you really need extra processing power, it is best to run it in normal mode.
It is also important to remember that Python is not the first choice for devices to achieve excellent performance. Both devices support the C++ API, which is also my “trick” to get the best performance from the devices in my tests.
2. Software Support
NCS2 can be used with Ubuntu, CentOS, Windows 10 and other operating systems. It can support TensorFlow, Caffe, ApacheMXNet, PyTorch and PaddlePadle through open neural network conversion.
CUA does not support Windows, but can run under Debian 6.0 or higher version (or any other derivative version, such as Ubuntu 10.0+). It is worth mentioning that CUA can only officially run the TensorFlow Lite model.
Application scenario:
Intel NCS2 also provides prototype, verification and deployment of DNN. For driverless and driverless vehicles and Internet of Things devices, low power consumption is essential. For those who want to develop deep learning inference applications, NCS2 is one of the most energy-efficient and lowest-cost USB sticks.
For example, the upcoming Titanium AIX also has an Intel Movidius Myraid X computing acceleration chip, which lowers the threshold for AI learning and development, helping AI enthusiasts and developers quickly build AI applications that can listen, speak, and watch.
Google Coral is more than hardware. It easily combines the functions of customized hardware, open software and advanced AI algorithms and provides high-quality AI solutions. Coral has many application cases in helping industrial development, including predictive maintenance, anomaly detection, robotics, machine vision, and speech recognition. It is of great application value to manufacturing, healthcare, retail, smart space, internal monitoring and transportation departments.
