Ambiq 展示 compressionKIT 开发套件

本站添加: 2026-05-23

[00:00:04]Hi, I'm Dr. Adam Paige, head of AI at Amvic. As more AI moves to the edge, efficient inference is only part of the problem. Devices like smartwatches, rings, and fitness bands generate continuous sensor data around the clock.

[00:00:18]And that data is still driving memory usage, wireless bandwidth, and battery life. At the same time, that data is extremely valuable. Teams need it to refine their algorithms, build longitudinal health insights, and deliver more personalized user experiences. Introducing Compression Kit. Compression Kit is our AI based codec technology for compressing sensor before it is stored, transmitted, or analyzed. It reduces ondevice memory, lowers transmit bandwidth, and opens up flexible inference paths from local to cloud. Today, I'll walk through a live PPG demo and show how teams building health wearables, hearables, and other sensing products can use compression kit to tune the balance between efficiency and signal fidelity.

[00:01:02]Here is the compression kit dashboard.

[00:01:05]This is a real-time compression and reconstruction workflow. Today's example uses PPG, which is one of the most common continuous signals in health wearables. The compression kit applies just as well to ECG, whether in a fitness watch or a disposable poultry patch, accelerometer data, and other sensor streams. On the left, we have the original input signal. Below it, the reconstructed signal after compression and decoding. You can compare the two waveforms side by side or switch to a delta view to see exactly where differences appear.

[00:01:37]On the right side is the control panel.

[00:01:39]It has two views you can page between.

[00:01:42]The first is model setup where developers select the model family, compression rate, and input record. The second is noise controls, which lets you inject live noise perturbations like baseline wonder, transient injection, and Gaussian noise directly into the input signal. I'll come back to that in a moment. The compression rate is the key dial here. It ranges from 2x all the way up to 16x, so teams can choose the operating point that fits their product.

[00:02:09]Let me start at 16x to show the most dramatic result. At 16x compression, we reach roughly 20x dynamic compression and 95% bandwidth savings. The air rate is around 11% PRD and yet the core signal structure is still clearly preserved. For products where battery life is the top constraint, this kind of reduction in wireless transmission can translate directly into longer time between charges.

[00:02:34]For comparison, I'll briefly switch to 4X. Here the error rate drops to 6.1%.

[00:02:39]Dynamic compression is about 5x. Memory saving reaches 50%. And bandwidth savings hits 80%. And at 2x the reconstruction signal is nearly indistinguishable from the original with only 4.8% error. So the full range from 2x to 16x is available and the dashboard makes it easy to evaluate each one.

[00:03:02]Now let me show the noise control panel.

[00:03:04]I'll switch the noise control panel to noise yield. You can see sliders for baseline wonder, transient injection, and galaxy noise. These let you inject real world noise conditions directly into light input signal. Watch what happens as I push baseline wonder and galsian noise all the way up.

[00:03:21]The original signal on top becomes visibly noisier, but look at the reconstruction below. The model is not just compressing, it is also cleaning up the signal. Because this is an AI based model trained across different noise conditions, it has learned to separate meaningful signal structure from noise.

[00:03:36]So compression and denoising happen together as part of the same pipeline.

[00:03:41]This is a significant advantage over traditional codecs which would faithfully compress the noise along with the signal. Here the models learn what matters and it preserves it.

[00:03:51]Notice how the compression intelligence panel updates live as they change settings. The error rate measures reconstruction quality, compression ratio, memory saved, and bandwidth saved. Translate that directly into system level impact. Less memory, less radio time, longer battery life. This also creates real flexibility in how AI gets deployed. Some teams will want to infer locally on compressed representations. Others will want to upload compressed data for cloud-based analysis, algorithm refinement, AB testing against ondevice models.

[00:04:22]compression kit supports both pass and the dashboard lets you evaluate either one. The dashboard also supports live streaming from any hardware over USB.

[00:04:31]The encoder runs in 4.1 milliseconds using 31.7 microjles per inference and fits in a 21 kilob memory footprint.

[00:04:38]This is not just a visualization tool is a direct path from ondevice deployment to real-time evaluation.

[00:04:45]The goal here is to make compression tangible. Developers can see exactly what they gain in efficiency, what they preserve in signal quality, and how each operating point fits their product.

[00:04:54]Compression kit gives developers a practical way to handle continuous sensor data more efficiently at the edge, reducing memory, bandwidth, and power while preserving the signal information that matters. Whether you're building a smartwatch, a health ring, a fitness band, or any device with continuous sensing, Compression Kit creates the path to longer battery life, more efficient data handling, and more flexible AI architectures. Compression Kit is currently in beta, and we're actively looking for early partners and customer feedback to shape what comes next. If you want to learn more or try Compression Kit with your own sensor data, reach out to us at Ambic. Thank you for watching.