Today, YOU learn how to put AI on FPGA.

Рет қаралды 7,340

BRH - French SoC Enjoyer

Күн бұрын

Пікірлер: 61

@CourageToGroww 3 ай бұрын

I am not sure how having the details of the project make this as entertaining as possible

@nolanbayon69 3 ай бұрын

I think he meant that i won't go into details in the video, but leaving it for people who wants to try

@BRH_SoC 3 ай бұрын

Exactly, As stated on the outro, I'd like to focus more on "project reports" in the future, giving little to no details, but still giving ppl a nice github for them to do it on their own. Finding the balance is actually pretty hard haha

@SiontheRapadant Ай бұрын

Finally, a video that shows and provides resources, instead of one that just talks about FPGA ai acceleration or is a teleconference video. Really glad you made this, thank you so much

@BRH_SoC Ай бұрын

Thank you, very glad it can be useful :)

@coove3940 3 ай бұрын

My friend actually did kind of the same for his bachelor's diploma work. He wrote a neuronetwork that identifies a modulation mode for the inputted signal and then put it on a FPGA. I thought it wasn't that hard when he explained it to me, but after watching this video I changed my mind, It is quite hard. He won the local diploma work contest with this work by the way. Maybe I will try to do the same for my diploma work this year. Thanks for inspiration!

@BRH_SoC 3 ай бұрын

Good luck to you :)

@inkybz 3 ай бұрын

hey man im in my first semester for my EE degree, I plan to write a research paper and I'm laying out the framework right now. This is SUPER duper gonna help me as I do some parallel wicked type learning thing I do for myself while the landscape changes.

@BRH_SoC 3 ай бұрын

Good luck with your research ! Especially if you have do deal with backward propagation haha

@inkybz 3 ай бұрын

@@BRH_SoC ive got absolutely no idea but i hope within a couple of years ill have the understanding to get actually started with the paper hahahahah

@ahmadalissa9187 3 ай бұрын

Amazing job Good luck buddy🎉

@hammadsafeer4283 3 ай бұрын

love these videos, Keep making these

@BRH_SoC 3 ай бұрын

Sure will !

@mattmmilli8287 3 ай бұрын

Well done! Cool to have a follow up to your last vid. I hope your channel goes places! 🚀

@BRH_SoC 3 ай бұрын

Cheers mate ! 🚀

@LuisRoel 2 ай бұрын

Bro this is sick, great job!

@BRH_SoC 2 ай бұрын

Thanks bro

@opkp 3 ай бұрын

great video, loved it!

@4lexxxizz Ай бұрын

Best YT chanel ever. 😮

@BRH_SoC Ай бұрын

Thanks men ;)

@thejcedeno 3 ай бұрын

Great content, don’t try to be employable, keep it up with this content

@neuralthink Ай бұрын

We love it!

@devansh3053 2 ай бұрын

Hi ive been following along with this project and am super interested I actually want to do this as a project however I'm curious as in the initial video you mentioned you wanted to create a full 10hr course on how to do this project is the github tutorial the course you were mentioning? i would love to hear back from you and FYI is this a super cool project

@BRH_SoC 2 ай бұрын

The course is open source, it is pinned on my gihub profile (with all of the slides). However, the notebooks are not as detailed and the slides may be tricky to understand without a lecturer. If you have any trouble, don't hesitate to check the learning resources on the github or use the comments section on this post on my blog : 0bab1.github.io/BRH/posts/PY2FPGA/ (because these actually sends me notifications and are easier to work with for troubleshooting.) PS : the blog is something I'm actively working on right now, you are officially the first beta tester / early access member ;)

@harryli9422 Ай бұрын

Hello, I am currently doing the same project. But I am trying to perform object detection tasks with a quantized YOLO model, and create the stich IP so i can integrate it into the KV260 KRIA FPGA board as the deep learning unit to run inference and output it via HDMI. But I seem to get stuck on the out of context synth, another problem is when i ran synthesis on the stiched IP in vivado, it just kills me VM OS due to limited resources. I used 2024.1 version of the XIlinx tool chain, do you think this will be an issue? Since I saw you use 2023.2 verison, but the FINN documentation recommended the 2022.2 version.

@BRH_SoC Ай бұрын

Hello, great project ! I never tried models as complex as yolo but a general recommendation is to stick with older versions of vivado (from 1-2 years ago) as the most recent ones tends to have unknown "features" (bugs) and older ones tend to have better support. I also know using a VM can limit vivado sometimes. The best is have a real machine for that but I don't know how much of a difference that can make. Hope this helped, do you have any specific error ?

@harryli9422 Ай бұрын

@@BRH_SoC Thank you for the suggestions~!\

@em3755 2 ай бұрын

IN the Github it says theres a 9 hr course. Where can I buy it?

@BRH_SoC 2 ай бұрын

Hello, the course is open source, pinned on my github profile. But it was meant to be teached live so I don't know if you'll get more infos there but the slides might help. You also have this blog post 0bab1.github.io/BRH/posts/PY2FPGA/ that might help. I saw your other comment. I am busy right now working on the next video. But I can always try to help. Where can we get in touch ?

@somethingelse3381 2 ай бұрын

can you recommend other youtube channels that make similar content for fpgas/soc/hardware? By the way great vid!

@BRH_SoC Ай бұрын

Thank you ! Frankly I started doing these videos because it was pretty much hard to find. As of right now, I don't have many recommendations if not Phil's lab (KZbin) and Adam Taylor (internet blog) as reliable information sources

@anonimus5800 2 ай бұрын

Can you do a one with quartus? 🥺

@BRH_SoC 2 ай бұрын

Hey, maybe you can use the packaged IP in quartus ? I never really left the Xilinx thing when it comes to using FPGAs. I'll let you try that, tell me how it goes ;)

@sampratt8584 26 күн бұрын

Hey I keep getting issues of the DMA transfer failing and I honestly do not know why. It appears DMA is halted.

@BRH_SoC 25 күн бұрын

Hey, I had DMA halting too when something in the DMA's IP config was not right. Make sure scatter gather is disabled, that unaligned transfers are allowed etc...

@sampratt8584 25 күн бұрын

@@BRH_SoC I double checked my DMA with the one in the github and they match. I might just redo the whole process and see if that fixes it. However, I am having issues with vitis not recognizing any of the header files in main.c as well as issues with it running ps7_init. I am not sure if this could be causing the issue.

@em3755 2 ай бұрын

Can you help me get this working for my board 1-on-1? I'll pay

@BRH_SoC 2 ай бұрын

I replied to your other comment on the tutorial to answer your questiobs

@BRH_SoC 2 ай бұрын

Questions *

@emiliomartineziii2980 2 ай бұрын

@@BRH_SoCThats good, though if you are still a freelancer who wants some easy money (Doctor salary for 5 hours), feel free to let me know, since i'm on a time crunch to get this done. I do have to say that after 2 weeks, I wont need any assistance since the deadline for my research on this assignment will be over.

@emiliomartineziii2980 2 ай бұрын

@@BRH_SoC Also please stay active! I have many questions! Even if you cant immediately do a paid 1-1 live session, just asking you questions here will save me TONS of time. Like for example, the end goal of the first half was to simply generate three different onnix files right? Though only 1 is used?

@emiliomartineziii2980 2 ай бұрын

@@BRH_SoC Also, if your still a free-lancer, lmk becuase I am 100% willing to give doctor level rates for training since I am on a time crunch to get this done for school

@BILDEK 3 ай бұрын

how much took time the training process . Is it less that regular systems?

@BRH_SoC 3 ай бұрын

I talked briefly about it in the video. Training time is longer when using quantized aware training. In the backend, the computer simulates quant error in order to let the model "adapt". Alongside more complex backward propagation, it requires more time indeed. You can also train normally and use Post Training Quantization, which works but the model is not adapted leading to greater loss. It all depends on your requirements, hope it helped ;)

@premixmpk7070 2 ай бұрын

Win Win

@АлябьевКонстантин Ай бұрын

Nah bro, I think it's better to run deep in technology to make it more like a tutorial, case now It looks like Mr.Beast show. The theme is truly interesting, but if it interesting for me I would like to recreate it. And moreover you've used a lot of stuff but didn't explain why you've chosen this particular technological stack. Anyway, great video! Thank you!

@BRH_SoC Ай бұрын

Thanks for the actually constructive feedback. What I aim to do is to introduce the subject in the video and then give resources (e.g. the github) for the people that want to actually build the thing. I feel like if I go too deep in the nerdy detail people will just pass onto other subject. I know I can also do some format where I really go in the details in a way that is very captivating (like 3blue1brown or mentalOutlaw on YT) but these styles asks for a lot of creativity/qualities/skills/background I think I lack for now. I'll try to mix styles a little bit on the next one : "brainrot" at the beginning to make the subject look dynamic and interesting and then add details to have a good overview of the subject. Anyways, thanks for the feedback, have a good day.

@АлябьевКонстантин Ай бұрын

@@BRH_SoC What a legend! Just don't take my words personally. I prefer neardy stuff cause I already in the field. Your github tutorial is very promising. If you still interested in AI on FPGA tutorials I can advise Marco Winzker tutorials www.youtube.com/@marcowinzker3682 They are good but much more "academical" I guess

@АлябьевКонстантин Ай бұрын

@@BRH_SoC What a legend! Just don't take my words personally. I prefer neardy stuff cause I already in the field. Your github tutorial is very promising. If you still interested in AI on FPGA tutorials I can advise Marco Winzker tutorials which is about low level implementation www.youtube.com/@marcowinzker3682 They are good but much more "academical" I guess