Good job of putting this board in an appropriate context. I'm a long time Pi user that purchased a Jetson Nano. The Nano was a good learning experience but I was reluctant to use it in my "edge" AI applications. Too expensive, too big, generates too much heat. I wanted to put it outdoors but active fan cooling would be a problem. I went to buy a Maxx II but as the other commenters said, "sold out". I'm tired of listening to reviews of unobtainable hardware like the Pi 4 compute module. I don't plan on joining any waiting lists.
I don't think it's going to be as bad as Pi 4 CM - I know there is stock available in Sipeed China Mainland taobao shop. And I already told our PM and logistics department to sort the sold out issue. Well, for Jetson Nano - I have to say, Nvidia came a long way, from basically unusable Jetson TK1 (I think that was their first SBC) to the most recent Jetson Nano 2 Gb - which at least has an attractive price for CUDA-enabled SBC, but still suffers from other problems you mentioned: heat, size and power consumption. Having said that, I think 2Gb version is worth the money, as a starting point to NVIDIA applications ecosystem (Isaac SDK, ISaac Gym, Transfer Learning Toolkit, etc).
I have two platforms based on kendryte k210, I hope their development and support by the manufacturer will not decrease with the advent of a new platform based on another soc. And yes, I already want to try maix 2 and talk about it on my channel.
Hi, I cannot find where to download .bin and .param files for the neural networks from the libmaix examples. Yolo2_20class_awnn.bin and Yolo2_20class_awnn.param. Also a video or article on converting models to awnn format would be very helpful.
They were uploaded just recently - here are the links: maixhub.com/modelInfo?modelId=26 maixhub.com/modelInfo?modelId=25 And yes, the model training/conversion tutorial is definitely in plans!
Спасибо за обзор и сравнение, с удовольствием смотрю ваш канал. Я имею почти все модели k210 bit, go, m1n, m1w , но поддержка Sipeed не очень хороша, более интересные проекты такие как asr или самообучение, имеют только демо примеры без фреймворков для их кастомизации. Maix-II является не продолжением линейки Kendryte k210 risc-v, а форком Allwinner V831 IP Camera SoC с коммерческим SDK, поэтому и тут вряд ли стоит ожидать хорошей поддержки. Maix-ii уже в продаже но с поддержкой уже проблемы, до сих пор не предоставлен инструмент для конвертации моделей, т.е на сегодняшний день это не больше чем игрушка.
Да, есть такая проблема, хотя Sipeed если сравнивать с другими китайскими производителями еще неплохо поддерживают продукты - у них ест рабочий форум и вики, плюс активный гитхаб - даже через 3 года после выпуска первых K210 плат. Они с января очень активно работали над Maix-II и другими новыми продуктами из серии Maix. По поводу конкертера - да, это большая проблема, я ожидал что будет опенсорсный, как nncase, но Allwinner даже бинарник предоставлять не хочет - в данный момент делается облачный сервис для конкертации.
Seems like everyone on Seeed is out of stock. They dont seem to be producing as much stuff as they used to or maybe this is because of the chip shortage, I cant tell.
Thank you for the information. I now own two of the M2_V381 docks. Could you please elaborate a bit more on the "Firmware flashing/Getting started" section? I have a design in mind. I am running windows 10 :( . It looks like VS Code can handle the development. Thank you in advance.
There's more details in the article. However I didn't have any luck with Windows version of flashing tool. Perhaps you can just install Ubuntu in VMWare VM?
@@Hardwareai Maybe i have used too many word-play in the second part, acronym fun - i hurry to explain what i meant - Risc-V is probably the first processor, representant of Open ISA (Instruction Set Architecture) - a factory can build a batch of RISC-V microprocessors - completely compatible with the binary code, not negotiating any ISA licences with companies that sell those for x86, ARM, ... MIPS. It will simply make devices like MAix cheaper and more widespread.
Hi, i found your channel very interesting there is a new Sipeed M1S board which is similar to M1 dock and maix ll devboards, i can't find any video tutorial or reviews for this board, if it is possible please give review on M1S devboard.
I would like to have one for my educational robotic project. Kudos for attached microphone. I wonder how it would work with offline speech recognition.
Depends on how open domain the task is. Surely it will work well for hot word detection and voice commands. For open domain speech transcription, judging by the fact it can handle at least 10 Mb models, I would say that is the reason to be optimistic. There is a cool demo for K210, where speech recognition model fits in about 2 Mb, www.bilibili.com/video/BV1C5411L7JC/ maixpy.sipeed.com/zh/course/speech/recognizer_cnn.html
Excelent video. This week my new Maix-ii board arrived, I have installed it, and the examples work, I agree with you it has better image quality than Maix-i , but I can't find documentation to convert my own keras or Tensorflow networks to the two files used by Maixpy3 (.bin and .params) , Could you share a link with the documentation for creating custom networks?. The Maixpy3 official WebSite is incomplete and the training chapter is empty.
Right... I need to find the time to make a tutorial on programming/converting models for MAIX II. In short, there is online model converter available at www.maixhub.com/modelConvert .
I remember working experience of open wrt with Intel Edison. I have 3 questions: # Can you compare the performance of RPi 4 + coral USB TPU performance vs MIAX 2 performance? # Can we run Opencv examples smoothly on MIAX2 with openwrt? # Can we work with the binocular camera module with MIAX 2 for depth and distance estimation?
All the good questions! #1 I won't be making a direct comparison, since it would be quite niche material. However, you can get ballpark estimation knowing that Edge TPU does 4 TOPS for inference workload and MAIX-2 does 0.2 TOPS. However RPI4 + USB TPU combination is horrible from price standpoint (95 USD) and power consumption (between 1000 mAh ~ 1500 mAh). #2 Not yet. I'll have more information about OpenCV in the next video. #3 I'll ask the PM in Sipeed about binocular module. I know the old binocular module for MAIX-I didn't have a global shutter, so there was no way (without attaching external hardware, e.g. FPGA) to receive picture from both cameras simultaneously.
@@Hardwareai Thank you so much for your nice replies. Did you check this K210 dock from Xalogic? Any comments on that. Because it would be really great resource. Frankly speaking K210 modules were nice with limitations. For example if I want to process voice assistant on edge + Object detection, no choice on K210 boards. If I want to do some posenet stuffs on k210, can not do. So after trying some object detection and classification stuffs on K210, I switched back again to RPi with tpu for general applications including opencv stuffs. www.hackster.io/news/xalogic-k210-ai-accelerator-looks-to-bring-kendryte-k210-powered-ai-acceleration-to-the-raspberry-pi-d7d4dcd3bf62 could solve the issue. Or any update Sipeeed wants to share with us? Thanks in advance
Right. For ASR + CV, K210 is really unsuitable, because of low RAM available. You could however train and run a MobileNet backbone Openpose, similar to what was done by OpenCV AI kit team for their board github.com/luxonis/depthai/blob/main/depthai_helpers/openpose2_handler.py this approach should work on K210 as well. For interfacing R Pi with K210, I think the main hurdle is software layer. Seeed studio had a Raspberry Pi hat with K210 www.seeedstudio.com/Grove-AI-HAT-for-Edge-Computing-p-4026.html pretty similar to Xlogic board you mentioned. The thing is, for the best performance and ease-of-use the K210 chip should be used in manner similar to Edge TPU with R Pi - data goes from SBC to K210, gets processed there and results are going back to SBC - all over high-speed SPI interface. That never was realized in Grove AI hat though, it was only using low-speed(UART) interface to get the results from K210, which was getting data from on-board camera. Another note aside - the price for Xlogic board is horrible :) I thought Grove AI hat was expensive
Buenos dias, How can i train my own model for the maix II. i´m working with keras and with pytorch, and I don´t know how to generate the .bin and .params files. Thank you.
Buenos dias! Well, the .bin and .params are generated with the help of the online converter. As I mentioned before I definitely will make a tutorial video on that topic, hopefully sooner than later xD
It's a pity they went with arm and not riscv, but if it can give it a more powerful som I won't complain. Will it support .tflite? Because k210, due to the limited resources, wasn't reliable
No, most of the NN accelerators don't support vanilla .tflite, just because they have specific optimizations. Even Google Edge TPU needs a conversion step. Currently MAIX II has a proprietary converter, but Sipeed is working on an open-source one.
Don't bother with the ' usual resellers ' as eBay and Amazon don't have them ( try with a VPN to see if outside North America gets any results that are not the Sipeed GO boards )
That would be a great project! I was thinking about using one of these cheap model fixed wing airplanes (less than 30 USD) and put a K210 module on it, then train a basic model that would allow it to stay in the air, akin to Donkey car.
Not worth the comparison really. Portenta's main processor is the dual core STM32H747 including a Cortex® M7 running at 480 MHz and a Cortex® M4 running at 240 MHz. So, Portenta is MCU based and MAIX II is CPU based. MAIX II is miles ahead in CV applications, especially considering it has a dedicated NPU. Although development might be easier for Portenta, since MAIX II is (partially) closed-source and sparsely documented as of now.
Hmmm, I'll give a response that I often give to open ended questions like this: Many things are possible, it is just a question of how much efforts and time you're willing to put into that. To my knowledge there is no Mediapipe ports to this chip. Of course you can try compiling from source. However I think it will be rather slow.
...pitty. It's a rather capable little platforn. I estimate that it has become unavailable to western developers because of its potential military applications, by order of the CCP.
No need to go that far xD it was a chip shortage. The good news is - I visited Sipeed just last week and they have a new board / SoM, that is similar in price range, but will be available long-term.
I think the board never really took off, because in the beginning it was out of stock a lot due to chip shortages. So the company did not have much incentive to develop the software. Btw, where you have been looking for docs?
Good job of putting this board in an appropriate context. I'm a long time Pi user that purchased a Jetson Nano. The Nano was a good learning experience but I was reluctant to use it in my "edge" AI applications. Too expensive, too big, generates too much heat. I wanted to put it outdoors but active fan cooling would be a problem.
I went to buy a Maxx II but as the other commenters said, "sold out". I'm tired of listening to reviews of unobtainable hardware like the Pi 4 compute module. I don't plan on joining any waiting lists.
I don't think it's going to be as bad as Pi 4 CM - I know there is stock available in Sipeed China Mainland taobao shop. And I already told our PM and logistics department to sort the sold out issue.
Well, for Jetson Nano - I have to say, Nvidia came a long way, from basically unusable Jetson TK1 (I think that was their first SBC) to the most recent Jetson Nano 2 Gb - which at least has an attractive price for CUDA-enabled SBC, but still suffers from other problems you mentioned: heat, size and power consumption. Having said that, I think 2Gb version is worth the money, as a starting point to NVIDIA applications ecosystem (Isaac SDK, ISaac Gym, Transfer Learning Toolkit, etc).
Wow, what an interwsting board. I can imagine cheap RC cars being sold now with a "follow item" mode if they include this board.
I think for that RC car to stay "cheap" it still needs a cheaper solution for CPU - mostly because from what I know retailer mark-up is crazy...
I have two platforms based on kendryte k210, I hope their development and support by the manufacturer will not decrease with the advent of a new platform based on another soc. And yes, I already want to try maix 2 and talk about it on my channel.
Yes, that's my hope too! K210 is still useful for many applications after all. You can order MAIX II - it's back in stock
Hi, please can you recommend me some tutorial for beginners, I have the board but I don't found a get started tutorial
You know, I meant to do the tutorial myself for quite some time, but other things push it out of the priority queue...
Hi, I cannot find where to download .bin and .param files for the neural networks from the libmaix examples. Yolo2_20class_awnn.bin and Yolo2_20class_awnn.param. Also a video or article on converting models to awnn format would be very helpful.
They were uploaded just recently - here are the links:
maixhub.com/modelInfo?modelId=26
maixhub.com/modelInfo?modelId=25
And yes, the model training/conversion tutorial is definitely in plans!
Excelent info, thanks!
Glad it was helpful!
I want to find a mic array app for one of these Sispeed devices but cannot find anything. Can you help?
It's Sipeed :) you mean for MAIX II?
Спасибо за обзор и сравнение, с удовольствием смотрю ваш канал. Я имею почти все модели k210 bit, go, m1n, m1w , но поддержка Sipeed не очень хороша, более интересные проекты такие как asr или самообучение, имеют только демо примеры без фреймворков для их кастомизации. Maix-II является не продолжением линейки Kendryte k210 risc-v, а форком Allwinner V831 IP Camera SoC с коммерческим SDK, поэтому и тут вряд ли стоит ожидать хорошей поддержки. Maix-ii уже в продаже но с поддержкой уже проблемы, до сих пор не предоставлен инструмент для конвертации моделей, т.е на сегодняшний день это не больше чем игрушка.
Да, есть такая проблема, хотя Sipeed если сравнивать с другими китайскими производителями еще неплохо поддерживают продукты - у них ест рабочий форум и вики, плюс активный гитхаб - даже через 3 года после выпуска первых K210 плат. Они с января очень активно работали над Maix-II и другими новыми продуктами из серии Maix. По поводу конкертера - да, это большая проблема, я ожидал что будет опенсорсный, как nncase, но Allwinner даже бинарник предоставлять не хочет - в данный момент делается облачный сервис для конкертации.
Seems like everyone on Seeed is out of stock. They dont seem to be producing as much stuff as they used to or maybe this is because of the chip shortage, I cant tell.
Chip shortages
Why this chip is not popular and well documented? Any ideas it looks interesting.
There is quite some documentation, but it is mostly in Chinese :) Google translate will help you.
Does the current software support video encoding using H.264 to SD-Card for the MAIX-II?
There is simple encode mp4 demo (c code), but not yet ported to python yet. It most certainly will be though.
@@Hardwareai Is the SDK available? If so github?
Thank you for the information. I now own two of the M2_V381 docks. Could you please elaborate a bit more on the "Firmware flashing/Getting started" section? I have a design in mind. I am running windows 10 :( . It looks like VS Code can handle the development. Thank you in advance.
There's more details in the article. However I didn't have any luck with Windows version of flashing tool. Perhaps you can just install Ubuntu in VMWare VM?
@@Hardwareai I managed to load it with balenaEtcher. I almost tried to do it with the Jetson nano, but, Etcher did the job. Thank you for the reply.
Very good 👍
Thank you! Cheers!
Yes! Open ISA! Isa is older but better than PCI :), still used in industrial applications.
I honestly didn't understand how your comment relates to the video, sorry xD
@@Hardwareai Maybe i have used too many word-play in the second part, acronym fun - i hurry to explain what i meant - Risc-V is probably the first processor, representant of Open ISA (Instruction Set Architecture) - a factory can build a batch of RISC-V microprocessors - completely compatible with the binary code, not negotiating any ISA licences with companies that sell those for x86, ARM, ... MIPS.
It will simply make devices like MAix cheaper and more widespread.
Awaiting for the binocular module support.
See my comment above, about binocular module.
Hi, i found your channel very interesting there is a new Sipeed M1S board which is similar to M1 dock and maix ll devboards, i can't find any video tutorial or reviews for this board, if it is possible please give review on M1S devboard.
I'll ask Sipeed if I can get a sample!
I would like to have one for my educational robotic project. Kudos for attached microphone. I wonder how it would work with offline speech recognition.
Depends on how open domain the task is. Surely it will work well for hot word detection and voice commands. For open domain speech transcription, judging by the fact it can handle at least 10 Mb models, I would say that is the reason to be optimistic. There is a cool demo for K210, where speech recognition model fits in about 2 Mb, www.bilibili.com/video/BV1C5411L7JC/ maixpy.sipeed.com/zh/course/speech/recognizer_cnn.html
Hello, thank you for you interesting video, I would like to.know is is possible record videos in fhd without interfaces, only conected to power bank.
You mean with Maix II? Should be possible, the dev kit has an SD card.
Excelent video. This week my new Maix-ii board arrived, I have installed it, and the examples work, I agree with you it has better image quality than Maix-i , but I can't find documentation to convert my own keras or Tensorflow networks to the two files used by Maixpy3 (.bin and .params) , Could you share a link with the documentation for creating custom networks?. The Maixpy3 official WebSite is incomplete and the training chapter is empty.
Right... I need to find the time to make a tutorial on programming/converting models for MAIX II. In short, there is online model converter available at www.maixhub.com/modelConvert .
I remember working experience of open wrt with Intel Edison. I have 3 questions:
# Can you compare the performance of RPi 4 + coral USB TPU performance vs MIAX 2 performance?
# Can we run Opencv examples smoothly on MIAX2 with openwrt?
# Can we work with the binocular camera module with MIAX 2 for depth and distance estimation?
All the good questions!
#1 I won't be making a direct comparison, since it would be quite niche material. However, you can get ballpark estimation knowing that Edge TPU does 4 TOPS for inference workload and MAIX-2 does 0.2 TOPS. However RPI4 + USB TPU combination is horrible from price standpoint (95 USD) and power consumption (between 1000 mAh ~ 1500 mAh).
#2 Not yet. I'll have more information about OpenCV in the next video.
#3 I'll ask the PM in Sipeed about binocular module. I know the old binocular module for MAIX-I didn't have a global shutter, so there was no way (without attaching external hardware, e.g. FPGA) to receive picture from both cameras simultaneously.
@@Hardwareai Thank you so much for your nice replies. Did you check this K210 dock from Xalogic? Any comments on that. Because it would be really great resource. Frankly speaking K210 modules were nice with limitations. For example if I want to process voice assistant on edge + Object detection, no choice on K210 boards. If I want to do some posenet stuffs on k210, can not do. So after trying some object detection and classification stuffs on K210, I switched back again to RPi with tpu for general applications including opencv stuffs.
www.hackster.io/news/xalogic-k210-ai-accelerator-looks-to-bring-kendryte-k210-powered-ai-acceleration-to-the-raspberry-pi-d7d4dcd3bf62
could solve the issue. Or any update Sipeeed wants to share with us?
Thanks in advance
Right. For ASR + CV, K210 is really unsuitable, because of low RAM available. You could however train and run a MobileNet backbone Openpose, similar to what was done by OpenCV AI kit team for their board
github.com/luxonis/depthai/blob/main/depthai_helpers/openpose2_handler.py
this approach should work on K210 as well.
For interfacing R Pi with K210, I think the main hurdle is software layer. Seeed studio had a Raspberry Pi hat with K210
www.seeedstudio.com/Grove-AI-HAT-for-Edge-Computing-p-4026.html
pretty similar to Xlogic board you mentioned. The thing is, for the best performance and ease-of-use the K210 chip should be used in manner similar to Edge TPU with R Pi - data goes from SBC to K210, gets processed there and results are going back to SBC - all over high-speed SPI interface. That never was realized in Grove AI hat though, it was only using low-speed(UART) interface to get the results from K210, which was getting data from on-board camera.
Another note aside - the price for Xlogic board is horrible :) I thought Grove AI hat was expensive
Buenos dias, How can i train my own model for the maix II. i´m working with keras and with pytorch, and I don´t know how to generate the .bin and .params files. Thank you.
Buenos dias! Well, the .bin and .params are generated with the help of the online converter. As I mentioned before I definitely will make a tutorial video on that topic, hopefully sooner than later xD
Hi, do you know where is the training part (with rest18 for example) associated with the inference part demo ?
I don't think that example exists yet. You can try training plain ResNet18 and then convert it with online converter and see how it goes first?
@@Hardwareai Which online converter you are refereing ? Actually the pre-trained model is in "awnn" format... I don't know about it.
It's a pity they went with arm and not riscv, but if it can give it a more powerful som I won't complain.
Will it support .tflite? Because k210, due to the limited resources, wasn't reliable
No, most of the NN accelerators don't support vanilla .tflite, just because they have specific optimizations. Even Google Edge TPU needs a conversion step. Currently MAIX II has a proprietary converter, but Sipeed is working on an open-source one.
Sold out! 😭😭
Don't bother with the ' usual resellers ' as eBay and Amazon don't have them ( try with a VPN to see if outside North America gets any results that are not the Sipeed GO boards )
I'm asking our PM and logistics to solve this issue :) Hold on
You can pre-order now
Launch one on a model rocket
That would be a great project! I was thinking about using one of these cheap model fixed wing airplanes (less than 30 USD) and put a K210 module on it, then train a basic model that would allow it to stay in the air, akin to Donkey car.
Can you compare it to the Arduino Portenta with vision?
Not worth the comparison really. Portenta's main processor is the dual core STM32H747 including a Cortex® M7 running at 480 MHz and a Cortex® M4 running at 240 MHz.
So, Portenta is MCU based and MAIX II is CPU based. MAIX II is miles ahead in CV applications, especially considering it has a dedicated NPU. Although development might be easier for Portenta, since MAIX II is (partially) closed-source and sparsely documented as of now.
Hi, can we implement mediapipe on MIAX II?
Hmmm, I'll give a response that I often give to open ended questions like this: Many things are possible, it is just a question of how much efforts and time you're willing to put into that. To my knowledge there is no Mediapipe ports to this chip. Of course you can try compiling from source. However I think it will be rather slow.
@@Hardwareai thanks for your reply. I may try. If successful I will comment back here. Anyway thanks for your nice videos. I love your contents.
Fuera de stock :/
It's back to stock! :)
...pitty. It's a rather capable little platforn. I estimate that it has become unavailable to western developers because of its potential military applications, by order of the CCP.
No need to go that far xD
it was a chip shortage.
The good news is - I visited Sipeed just last week and they have a new board / SoM, that is similar in price range, but will be available long-term.
Seem to be badly supported and documented.
I think the board never really took off, because in the beginning it was out of stock a lot due to chip shortages. So the company did not have much incentive to develop the software.
Btw, where you have been looking for docs?