Getting Started with Edge AI Chips: From Boot to Offline Translation

I spent the weekend tinkering with the on-device features of a learning machine powered by the G100 high-efficiency SoC and ran into quite a few pitfalls.

Let me clarify first: The G100 is a chip from Qianhe Yibang, not to be confused with the Helio G100 found in smartphones. Regular users don't touch the chip itself; what they use are translation pens and learning machines equipped with it. It falls under the category of 3D compute-in-memory. Simply put, this means placing computation and data storage closer together to reduce data movement, saving power and speeding up startup. After my actual experience, I mainly focused on whether voice and image recognition could produce results when offline.

First, let's explain a few terms. An SoC is a small system that packs the CPU, memory controller, and AI computing onto a single chip; "on-device" (edge-side) means the device computes locally without sending data to the cloud first; a "model" refers to the local recognition algorithm, essentially an offline dictionary.

Preparation before starting. Find a terminal equipped with the G100 and keep the battery fully charged. Connect to WiFi first—don't rush to enable airplane mode, as many devices need to download models beforehand. Prepare a Chinese sentence like "The weather is nice today" and an English line like "The library is on the second floor." Place a printed sheet of paper on the desk and ensure the environment is as quiet as possible.

After powering on, long-press the power button to enter the main interface. Look for entries labeled Voice Translation or Photo Text Recognition, prioritizing offline functions. If a model download prompt appears, click Confirm Download. Wait until the progress bar finishes and you see the message Model Ready. Until then, the offline buttons might remain grayed out. Next, enable Offline Mode, then turn on airplane mode, expecting all computations to stay local.

During testing, tap Start Recording and say "The weather is nice today." Wait a few seconds; you should see the Chinese transcription and English translation. Expect the transcription to be mostly complete and the translation to be understandable. Then tap Take Photo, aim at the English text on the paper, press the shutter, and observe the recognized text and Chinese definition. Expect long sentences to potentially break lines, but core words shouldn't be wrong. Finally, tap Copy or Save to Notes to export the results.

The most common blocker is grayed-out entry points. I initially thought the device was broken, only to realize the model hadn't finished downloading. The solution is to connect to WiFi, avoid switching to the background, and wait for it to load completely.

Recording and taking photos are also prone to failure. If someone nearby is talking or you speak too fast, characters will be dropped in the transcription; if the paper reflects light or the angle is skewed, letters might be misidentified as numbers. Speak slower, keep sentences shorter, and shoot perpendicularly for higher success rates.

Another pitfall is having expectations that are too high. The G100's strengths lie in real-time voice processing, image handling, quick startup, and low power consumption. It is not suitable for acting as a cloud-based large model. Asking it to handle complex long documents or multi-turn conversations often leads to slowness and errors.

The conclusion depends on the situation. If your needs are looking up words offline, recognizing text via photo, or transcribing classroom recordings, it's worth trying. If you expect it to write articles or do mathematical proofs offline, don't hold your breath.

These small devices will increasingly rely on local models with cloud fallbacks, focusing on energy efficiency and speed at the edge. My next step is to take the same sentence and the same piece of paper, run them through this device and phone-based cloud translation separately, and see which one still works when disconnected.