Beaglebone Ai Heating Issues

Hi ipv1 ,

I  looked into this and this is a known issue:

BeagleBone AI - Frequently Asked Questions (FAQ)

They recommend adding a larger heat sink and cooling via fan.   

My investigations showed much the same. My questions are:

1. How was this not caught during testing ? Some minor overheating might be missed but shutting down as described on many of the comments makes it obvious that there is a serious problem.

2. This is happening when you are only running the OS on the ARM Cortex-A15 cores.  What happens to the temperature when you start using some mixture or all of the ARM Cortex-M4s cores,

      DSPs, EVEs and PRUs too?

3. If it needs a bigger heat sink and fan just running the OS why doesn't it HAVE a bigger heat sink and fan (see 2.).

After reading the earlier comments in this discussion and checking for other comments on the web I hopped over to Ebay and bought a couple of 25mm fans. The heat sink on the AI is 25mm and

it looks like there are threaded holes to attach the fan with. I'm not sure how you mount a larger heat sink  other than a adhesive thermal pad or thermal epoxy. There is an adhesive thermal pad

between the heat sink and SoC but I'm not a big fan of thermal pads and in this case I would much rather have a heat sink that uses mount points on the PCB and a wafer thin layer of quality thermal

paste.

It also looks like it could use some serious thermal management on the OS side. I'll give them the benefit of the doubt and hope that somebody is looking into this.

Surely you can change its frequency and governer to low and power saver respectively, but beaware in low power mode it might shut down few peripherals. Also it surely needs active cooling source, read thermal considerations for AM57xx http://www.ti.com/lit/an/sprac53a/sprac53a.pdf 
Better get proper cooling method to test it at full efficiency.
Caution don't let it get too hot, there are chances of EOS damage.  >> http://www.ti.com/lit/an/sprace7/sprace7.pdf

For any upcoming show maybe it could be better to use something more mature.. most of the Pi releases had software teething issues at launch, the software took a while to be in a mature state.

NVIDIA and BB-AI's chips are playing in different markets I think.. the chip in the BB-AI (AM572x) is designed for automotive applications, so will likely have the guaranteed X amount of years availability, and comes in an industrial temp version too, and might use more traditional methods of processing (multiple CPU cores and some general purpose accelerators in H/W, and some accelerators for image recognition features). NVIDIA's chip may not be available in many years time for use in customer designs (this is a guess though) if they keep churning out newer chips and want to sell those instead. NVIDIA has fantastic technology for doing state-of-the-art computing though.

The BeagleBone Black's AM3358 had the longevity advantage too, because that chip was (is) in production or at least offered for sale, for 10 years (might be 15, I can't recall).

The AM3358 was easy-ish to prototype with, to create custom designs (I tried to create my own too, using the OSD3358 which integrates that chip, PMIC, Flash and RAM), but the project went away - lots of time was wasted with Intel's Edison, we had a working prototype, and it went end-of-sale : (

Prototyping a custom board will be much harder with the BB-AI's chip, but there should be modules with that same chip I hope.

I believe development kits are basically to test processor functionality, to learn about new things - How it performs, its features usability, How to code them, most importantly to find mistakes/errors before using it for applications.. But the main thing is, they are made to be used for specific applications. They went good with previous boards similar to other board designer, but this is one mischievous processor. They just need to design the board properly.
This is just board designing gone bad.

I don't know how useful it is, but there's some Sitara Bootcamp, with Power Management training ppt : )

https://training.ti.com/sitara-arm-processors-boot-camp-linux-power-management-overview-and-hands

It's for AX335x, but the concepts mentioned in there are supported in the BB-AI/X15 (AM572x) chip too.

There's some Linux power management training link here:

Sitara Linux Training: Power Management - Texas Instruments Wiki

But it's using TI's Linux image, I don't know if the same drivers are integrated in the normal BB-AI Linux image yet, and if not, what the plan for that is.

The BeagleBoards are a bit more than development kits though.. there's some serious commitment to supporting them with software images and driver code, and supporting customers to create their own variants, or to integrate onto custom PCBs in their own designs, i.e. BeagleBone in appliances effectively, and these would be a large task if the Linux code/drivers/platform stuff needed to be maintained by the company creating their own design.

It's a known processor, it was the same chip on an earlier BeagleBone design, the X15. So they have several years of experience with this chip. It doesn't make sense to assume they had no idea it would get hot.

Both the BB AI and the Jetson Nano were released to target folks looking to get involved in AI and ML, so I am not seeing a diff in target market here.

Also, the Jetson Nano is Production ready.

Oh, please. I've been working with the BB for some time now and comparing the support from the NVIDIA site to the now BB Google forums, the NVIDIA support and community far outshines anything the BeagleBoard folks are doing.  If you can get a response from the BeagleBoard you are lucky.

And the fact the BB AI is using the same SoC as the X15 goes back to one of my original issues; its 6+ year old technology.  It was not designed for any of the newer AI and ML aps or workloads of today.

I've had great experiences with the BB team-members for years. They're human, everyone has their off days.

jomoenginer  wrote:

 

It was not designed for any of the newer AI and ML aps or workloads of today.

Ok, so what? There's new orgs being formed all the time, dealing with compute and AI at the edge. That's a growing deployment scenario, and even Cortex-M is being deployed for that role. The raw performance isn't the only criteria.. device longevity, industrial temp spec, hard real-time engines, are all things that the TI chips are good at.

It sounds like you expected BB-AI to be something different from what it is.

It's good that the Nano is guaranteed for sale until 2025.. I hadn't realized that.

https://developer.nvidia.com/embedded/community/lifecycle

the fan listed in the bom here: BB AI cooling Addon board available
is the one that fits on the heatsink that is mounted on the BBAI as default.

the fan listed in the bom here: BB AI cooling Addon board available
is the one that fits on the heatsink that is mounted on the BBAI as default.