We’ve written a fair bit about Google’s Cloud TPU, the AI inference and ML training accelerator the company developed for handling specific workloads more effectively than a GPU or CPU. These large-scale TPUs, however, are intended to power server rooms and major data centers. Google has also developed hardware for smaller devices, use-cases, and the network edge, appropriately dubbed the Edge TPU.
The Coral FAQ describes the Edge TPU as follows:
The Edge TPU is a small ASIC designed by Google that provides high-performance ML inferencing for low-power devices. For example, it can execute state-of-the-art mobile vision models such as MobileNet V2 at 100+ fps, in a power efficient manner.
The Coral development board includes:
- Edge TPU Module (SOM)
- NXP i.MX 8M SOC (Quad-core Cortex-A53, plus Cortex-M4F)
- Google Edge TPU ML accelerator coprocessor
- Vivante GC7000 GPU
- Cryptographic coprocessor
- Wi-Fi 2×2 MIMO (802.11b/g/n/ac 2.4/5GHz)
- Bluetooth 4.1
- 8GB eMMC
- 1GB LPDDR4
Various USB ports, audio connections, and video connections are all available, including multiple microphone hook-ups, USB-C support, an HDMI 2.0a port, and both a 39-pin and 24-pin connector for a MIPI-DSI display and a MIPI-CSI2 camera. The Edge TPU claims to support Mendel Linux (a derivative of Debian). I wasn’t able to find any information on this distro online. Google’s web page for the Edge TPU project only states that the device is compatible with “Linux OS.” This may be a typo in the documentation or a reference to a custom Linux distro that Google built itself. The dev board is on sale for $149.99.
Also up for sale is a new Edge USB accelerator. Unlike the complete development module, this is just an Edge TPU implemented via USB that can be connected to other devices, such as a Raspberry Pi. Think of it as conceptually similar to Intel’s Neural Compute Stick. Benchmarks for the Edge TPU and USB-connected version (according to Google’s FAQ) are below. The new USB accelerator is also cheaper than the full board, at $74.99:
The Edge TPU only supports TensorFlow Lite, so any code you want to run on the hardware must be converted to that format first. It can’t perform traditional backward propagation (necessary for traditional ML training), but there’s a modified way you can run these workloads in specific instances, as described in the FAQ.
For now, AI remains a fairly specialized field, and parts like this are intended for a relatively small developer audience that can use them for software testing. In the future, however, these types of low-cost devices could be key to extending AI capabilities to ‘regular folks’ and hardware, allowing more people to experiment with the kinds of features AI can provide.
Stadia Is Now Playable on iPhone Thanks to Google’s New Web App
Google promised iPhone support, but Apple's App Store policies got in the way. Now, there's finally a way to play Stadia on iOS — just fire up Safari and go to the Stadia site to use the new progressive web app.
Apple Hires Away Google’s Head of AI
Apple is gearing up for a fight with several high-profile AI hires. Its latest move is to poach John Giannandrea, Google's head of AI.
Google’s Pixel 3 Makes an Appearance in Open Source Code
As is tradition, the first hints of Google's next flagship smartphone have appeared in the Android open source code. Specifically, Google is optimizing code for a device called "Pixel 3." It doesn't get much clearer than that.
Google’s Cloud TPU Matches Volta in Machine Learning at Much Lower Prices
Google and Nvidia both offer competitive machine learning products — but Google is beating Nvidia on costs, at least in certain tests.