Okay. Thanks guys for the comments. I’ve used the murata module before in converting PIR sensors to Lora PIR sensors before. Not going to sleep properly I would most likely put down to software bugs in the library. That seems to have been the bug bear for a few libraries in the past.

I used a grasshopper board in the past but it’s ok longer made. This time I also need to receive.

I have no particular requirements. I’m expecting receive mode to draw around 30ma and deep sleep micro amps.

Did you managed to get micro amps with RP2350 in deep sleep? 

I am using an RP2040 as a sound recorder and I need to use an external RTC chip to shut it down until next audio sampling.

No. I was using an stm based chipset at the time.

Oh apparently it’s still in production 

Read "Grasshopper LoRaWAN Development Board from Tlera Corp on Tindie" on Tindie

Grasshopper LoRaWAN Development Board from Tlera Corp on Tindie

Grasshopper LoRaWAN Development Board from Tlera Corp on Tindie

Tindie

That sounds very challenging what you are trying to do with the recorder. I have a recorder project. My sound samples are around 1024 samples each. So around every 43ms or so I get one. You can’t miss any. I would be amazed if an external real time clock can help you out here. I would have thought that wake up on interrupt from what ever hardware is sending your sound data is more likely to be helpful. Also then you need to write the data away. So you might need a whole well functioning async library that supports deep sleeping in between times. That would be a very cool library.

If you manage to get this working I bet a lot of people would like to hear your story, me included.

Oh, I am afraid I was not explicit enough: I am powering off the rp2040 after it records some arbitrary length audio on SD card. I need to record audio hourly to collect data for a ML project so the recorder is nothing fancy. I am using and external DS3231 to rewrite the alarm register with a new alarm time then the DS3231 shut's down everything to be re-powered at a later time. RP2040 "deep sleep" is terrible, if I remember correctly I couldn't get under 10mA.

Now I am working to re-purpose that project to also analize the audio sounds recorded, send the results by LoRaWAN then sleep some abitrary time and so on.

I will look on your recorder project to understand it better and if you think the DS3231 library can help you I will put it on github.

Ah that clarifies things. I remember when I first started working on my lorawan project in 2018. Deep sleep was the problem for everyone. At that time the esp32 was looking very interesting for everyone. That also had troubles with deep sleep. I started using the stm based grasshopper which did do deep sleep properly. But each of these boards had different libraries. 

Sounds a bit like maybe the rp2350 might be at that stage. Though who knows. Is it certain that the other underlying hardware supports a deep sleep? I have no idea myself.

My recorder project is pi based. So real time clock waking up is well established. But my project has  uses much more  power of course. I think around 130mw when it’s running on a pi zero.

I read you Medium post and remembered I read it some time ago, wonderful!

Did you have explored this technology from Aspinity? 

Read "AML100-REF-1 - Aspinity" on Aspinity

AML100-REF-1 - Aspinity

The AML100-REF-1 is a fully integrated reference module for straightforward prototyping and testing using the AML100.

Aspinity • 16 Mar 2024

It's an ultra low power analogue ML chip which listens for specific sounds (glass breaking, gunshots, chainsaws) then if sound is detected it wakes up the digital companion mcu for further processing. Aspinity says it draws around 37 micro amps in always on mode which I believe is great for battery operated devices in remote locations.

Okay, now you got me excited! This is very interesting. Thank you very much for this link.

Good thread. We are exploring power conserving options like this. Thanks for the link @ioanF 

Any tips on reducing microSD card power consumption? Writing the data seems to be the most power consuming part. I know some cards are more efficient.

Unfortunately I haven't any ideas to reduce microSD card power consumption. I am struggling with audio ML and power consumption without greater success than anybody else. This struggles led me to Aspinity's analog chip. However they are understandable a little bit opaque, one needs to sign an NDA to get access to datasheets, programming framework, further information and development kits. Despite this I believe their technology is very promising for wildlife conservation, not as a bird song classifier for example, but at least to detect sound that are are easy to discriminate against the background sounds like gunshots, chainsaws etc.

I was also contemplating Georgia Tech FPAA technology, which is very similar with what Aspinity are doing, Anadigm or Okika Technologies, but it seems none of them are close to what Aspinity is offering.

Hmmm. I guess the question I would ask is did they mean “potentially” useful for those things but they haven’t done it yet.

Any model that can fit in such a small amount of space and use so little power is likely also going to have a lot of false positives I’m thinking. Then the question becomes how much can your use case tolerate ?

Good question... as usual with commercial applications is hard to really know if there are successful implementations. Also is frustrating that one can find such kind of answers after investing time, trust and money in their technology. I am still waiting their development kit to be ready to be delivered but very anxious if the commercial project I am working on will lead to achieve its goals.

However they claim their chips have far less false positives than TinyML models. Don't know yet but I think it is worth investigating.

Anyhow, the digital path to ML I believe is doomed to high power requirements.

I’m kinda with you about the higher power models.

Mainly because of my experience with smaller models with object detectors.

@ioanF , you can do a little bit better than 10 ma. I have here an adalogger rp2040 feather with DS3231 RTC wing and a I2S Mems microphone. During "dormant" mode running from xosc and waiting for DS3231 wakeup I get 4.7 ma . This includes about 1 ma for  microphone and DS323 together. OK, that is still 3 ma higher than 0.7 ma RP2040 documentation is said to claim. I guess, there is some uncertainty with  the AT35 USB tester I'm using. Putting the LDO enable pin (EN) to ground, the USB tester said 1 ma, which may be dominated by the offset of the tester as the LIPO charger as remaining component should only consume 0.2 ma (max).

Edit (for the records): After 'disabling' all GPIO but the RTC wakeup pin, the system (RP2040+DS3231 RTC+ I2S Mems microphone) consumes only 1.8 mA. I guess I`m now close to the clamed dormant power consumption!

Valuable information.  Thank you Akiba. I'm probably going to give up on the rp2350 at this stage it seems.

Akiba, have you used the MKR wan ones ? Do you know if the deep sleep works properly in combination with transmitting?

I haven’t worked with that flavor of Arduino board. I mainly stick to custom AVR and ARM boards that are application specific so have little experience with many off the shelf boards.