3 ms·

I think this is partly due to the fact that there is marked growth in interest in embedded systems. MCUs have gotten even cheaper. It's now standard to have mul

I think this is partly due to the fact that there is marked growth in interest in embedded systems. MCUs have gotten even cheaper. It's now standard to have multiple tiers of sleep modes (standby, sleep, deep sleep). Multiple clock rates are becoming more popular - which enables on-demand heavy inferencing interspersed among periods of lower processor usage. CMOS sensors increasingly have on-chip motion detection[1] that uses less energy than full sensing. RF comms are being managed more efficiently as for most embedded devices it is the largest use of energy by orders of magnitude. Last but definitely not least, machine learning at the edge is a nascent application as we realize on-edge inferencing can avoid the aforementioned energy drain of sending huge amounts of data to the cloud. Of course for mission-critical applications C is still the standard for firmware in the industry, but for many startups providing novel uses of edge computing and the plethora of hobby projects C++ is a solid choice. [1] https://www.ntu.edu.sg/home/eechenss/Papers/conf-2011-A%20CMOS%20Image%20Sensor%20with%20on-chip%20Motion%20Detection%20and%20Object%20Localization.pdf https://www.ntu.edu.sg/home/eechenss/Papers/conf-2011-A%20CM...

stbtrax 6y ago
Having been in emb systems for a decade+: I'm not sure I agree with that. MCU prices have been relatively stable for the past decade. Sleep modes etc have existed for a while and are not drawing in new people. VCs are investing less in HW startups. Most won't invest in ones that don't have recurring revenue. Imaging sensors are made by a few large players and CMOS on-chip motion is not that useful since you can do the same thing with PIR. AI accelerators coming in to emb linux class projects is a new thing but I'm not sure who is using that besides the big companies at this point. Also, the growth in C++ is probably in the new standards, which are not well supported for embedded compilers (emb linux excluded)
- swiley 6y ago
  IMO it's more the flash than the CPU that puts the lower bound on power consumption with the smaller MCUs once you start sleeping. >Also, the growth in C++ is probably in the new standards, This, especially since growth here means growth in queries.
  - datameta 6y ago
    > IMO it's more the flash than the CPU that puts the lower bound on power consumption with the smaller MCUs once you start sleeping. Agreed! There are some novel remarks on that in the link under "[1]" in my reply to the parent comment. Or see the following digest if paywalled: https://community.arm.com/developer/research/b/articles/posts/m0n0-an-arm-research-platform-for-n-zero-sensors https://community.arm.com/developer/research/b/articles/post...
- datameta 6y ago
  Perhaps the recent industry growth isn't as great as I outlined. However I do believe that many more small projects are being started up (one of the reasons could be as simple as newly found time during months of quarantine). I do maintain that CMOS on-chip motion is useful because adding a PIR sensor is prohibitive for some form factors and more importantly has a lower effective range. What I find even more interesting, from a power usage standpoint, is this recent research on Adaptive Video Subsampling for Energy-Efficient Object Detection[0]. This buys some space on the energy budget for running ML on the device. Although image sensor reading is energy intensive compared to read-write and processor ops by 3 orders of magnitude, as outlined in the second slide of [0] we can see that LTE comms are a further 3-5 orders of magnitude more expensive than image sensor reads. Topically, here is some research on real-time wake from ultra-low power sleep (~10nW) and lowered idle SRAM usage [1] as well as the proposal for no power IR sensing [2]. [0] https://www.tinyml.org/summit/posters/Jayasuriya,%20tinyML%20Summit%20poster.pdf https://www.tinyml.org/summit/posters/Jayasuriya,%20tinyML%2... [1] https://ieeexplore.ieee.org/document/9063136 https://ieeexplore.ieee.org/document/9063136 [2] https://eri-summit.darpa.mil/docs/ERIPoster_Applications_N-ZERO_Northeastern.pdf https://eri-summit.darpa.mil/docs/ERIPoster_Applications_N-Z...

stbtrax 6y ago

Having been in emb systems for a decade+: I'm not sure I agree with that. MCU prices have been relatively stable for the past decade. Sleep modes etc have existed for a while and are not drawing in new people. VCs are investing less in HW startups. Most won't invest in ones that don't have recurring revenue. Imaging sensors are made by a few large players and CMOS on-chip motion is not that useful since you can do the same thing with PIR. AI accelerators coming in to emb linux class projects is a new thing but I'm not sure who is using that besides the big companies at this point. Also, the growth in C++ is probably in the new standards, which are not well supported for embedded compilers (emb linux excluded)

swiley 6y ago
IMO it's more the flash than the CPU that puts the lower bound on power consumption with the smaller MCUs once you start sleeping. >Also, the growth in C++ is probably in the new standards, This, especially since growth here means growth in queries.
- datameta 6y ago
  > IMO it's more the flash than the CPU that puts the lower bound on power consumption with the smaller MCUs once you start sleeping. Agreed! There are some novel remarks on that in the link under "[1]" in my reply to the parent comment. Or see the following digest if paywalled: https://community.arm.com/developer/research/b/articles/posts/m0n0-an-arm-research-platform-for-n-zero-sensors https://community.arm.com/developer/research/b/articles/post...
datameta 6y ago
Perhaps the recent industry growth isn't as great as I outlined. However I do believe that many more small projects are being started up (one of the reasons could be as simple as newly found time during months of quarantine). I do maintain that CMOS on-chip motion is useful because adding a PIR sensor is prohibitive for some form factors and more importantly has a lower effective range. What I find even more interesting, from a power usage standpoint, is this recent research on Adaptive Video Subsampling for Energy-Efficient Object Detection[0]. This buys some space on the energy budget for running ML on the device. Although image sensor reading is energy intensive compared to read-write and processor ops by 3 orders of magnitude, as outlined in the second slide of [0] we can see that LTE comms are a further 3-5 orders of magnitude more expensive than image sensor reads. Topically, here is some research on real-time wake from ultra-low power sleep (~10nW) and lowered idle SRAM usage [1] as well as the proposal for no power IR sensing [2]. [0] https://www.tinyml.org/summit/posters/Jayasuriya,%20tinyML%20Summit%20poster.pdf https://www.tinyml.org/summit/posters/Jayasuriya,%20tinyML%2... [1] https://ieeexplore.ieee.org/document/9063136 https://ieeexplore.ieee.org/document/9063136 [2] https://eri-summit.darpa.mil/docs/ERIPoster_Applications_N-ZERO_Northeastern.pdf https://eri-summit.darpa.mil/docs/ERIPoster_Applications_N-Z...