[kkielhofner]

I’ve worked with Jetson going back to the TK1 and I highly recommend you do not do this.

Nvidia has significant dominance in the AI space because of their work on software and the overall platform.

With the Jetson line being the sole exception. Use it for what it’s for - a targeted build for an embedded/specific application requiring small size and low power.

The software is a mess. Support for Jetson (generally) is a far afterthought or not considered at all around projects at Nvidia and the broader ecosystem. When it is supported at all it lags behind significantly, using ancient distros (Jetpack), etc. To make matters worse the user base is so (relatively) tiny there are bugs and strange behavior everywhere.

Just don’t do it.

[dchichkov]

This is a bit surprising to hear. Current Jetpack 6 is Ubuntu 22.04 - this is the current Ubuntu LTS release. There's nothing ancient about it, no? I'm pretty sure, if I go and check versions of CUDA, PyTorch, Tensorflow - it'd be also relatively recent.

I'd suggest checking what examples are available, see what community is doing, see if what you need had already been tried - https://www.jetson-ai-lab.com

From what I've seen, mainstream LLM libraries like VLLM, llamacpp that use CUDA under the hood tend to work out-of-the-box. And there are tutorials available: https://www.jetson-ai-lab.com/tutorial_text-generation.html. I think that TensorFlow/Pytorch are also well maintained, although I've not checked recently.

[kkielhofner]

I think this perspective comes from a lack of historical experience and hands-on experience overall.

Nvidia more broadly has very impressive support for their GPUs. If you look at the support lifecycles for their Jetson hardware over time it's significantly worse. I encourage you to look at what support lifecycles have looked like, with the most "egregious" example being dropping of support for the Jetson Nano in from what I recall was within a couple of years.

Another consideration - Jetson is optimized for power efficiency/form-factor and on a per $ basis CUDA performance is terrible. The power efficiency and form-factor come at significant cost. See this discussion from one of my projects[0]. I evaluated the use of WIS on an Orin Nano that I have and it was nearly 10x slower than a GTX 1070 which is seven years old and is still supported by the latest drivers and CUDA 12 on whatever OS you want.

Nvidia knows what they're doing in terms of productization and the Jetson line should not be seen as some kind of secret hack/unlock for getting CUDA performance with gobs of RAM. In the case of LLMs I wouldn't be surprised at all if CPU beats it and at that point pickup 256GB of RAM or whatever for equivalent cost.

In the end what do I care what people use, I'm offering the perspective and experience of someone who has actually used the Jetson line for many years and frequently struggled with all of these issues and more.

[0] - https://github.com/toverainc/willow-inference-server/discuss...

[dchichkov]

I've used Jetson for a few projects as a hobby. Made an I2S Sodar array with a TX2. And some robotics projects with a Jetson AGX Xavier that I got to evaluate and then to work on. And a few both, professional and toy projects with versions of Jetson Nano kit and Xavier. But this was between 2017 and 2021 or so.

About a year back, I took that very early version of AGX Xavier, that got released years ago. It wasn't even the version that was officially released. Yet I was able to refresh it to newer Ubuntu without any issues.

Wheels are often not pre-built for aarch64, yes. If you want to compile directly on Nano, disk performance is very important. Sometimes you get I/O bound.

Orin Nano being that slow in [0], it looks like you've been trying it in Aug 2023. It maybe worth re-evaluating on the latest Jetpack, it had transitioned to CUDA 12.2, TensorRT 8.6, cuDNN 8.9. I would expect that recent popularity of ASR/TTS pipelines and LLMs was not completely missed by Jetpack maintainers (there are some tutorials here - https://www.jetson-ai-lab.com ). And recently released JetPack could be optimized a lot more for these workflows.

And your project is very cool! I'd suggest sharing it and your performance numbers (!) with the maintainers of: https://developer.nvidia.com/embedded/community/jetson-proje...

[kkielhofner]

> I've used Jetson for a few projects as a hobby. Made an I2S Sodar array with a TX2. And some robotics projects with a Jetson AGX Xavier that I got to evaluate and then to work on. And a few both, professional and toy projects with versions of Jetson Nano kit and Xavier. But this was between 2017 and 2021 or so.

Nice! I'm sorry if I seemed dismissive or even disrespectful, in my experience Jetson certainly has it's place (why I've been using them for years) but compared to "bring your distro, apt-get/.run Nvidia driver" they can be a serious shock for casual users. Then they see the performance...

> Orin Nano being that slow in [0], it looks like you've been trying it in Aug 2023. It maybe worth re-evaluating on the latest Jetpack, it had transitioned to CUDA 12.2, TensorRT 8.6, cuDNN 8.9. I would expect that recent popularity of ASR/TTS pipelines and LLMs was not completely missed by Jetpack maintainers (there are some tutorials here - https://www.jetson-ai-lab.com ). And recently released JetPack could be optimized a lot more for these workflows.

Interestingly WIS was recently bumped to CUDA 12.2, etc and the performance improvements were very marginal. WIS uses Ctranslate2 under the hood (same as faster-whisper) which offers among the best Whisper performance overall but doesn't benefit much from changes in these underlying libraries. In the end even if it somehow magically doubled performance (it doesn't and won't) that still places the latest generation ~$600 Jetson board 5x slower than an ancient yet still fully officially supported ~$100 GPU. Power and form-factor is an issue but for the voice assistant use case a Jetson board barely doing realtime with Whisper medium is unacceptable to me and the vast majority of our users. Our goal is sub one second voice command sessions from end of speech, to command execution, to TTS response and Jetson just can't provide that at any cost.

I'm glad there are community resources for Jetson platforms (which I'm aware of) but their existence underscores my point - you'll notice when perusing through there are often various hoops to jump through whereas anything else is basically "install driver, container toolkit, docker run" and it just works and works performantly. Basically CUDA x86_64 and discrete GPUs is native/expected/developed for, Jetson is almost always a bit of an edge case with rough edges (relatively) all over the place.

> And your project is very cool!

Thanks! In terms of your suggestion I certainly might but in the meantime, overall (based on my Jetson experience) as I said in that discussion I'm very reluctant to officially support the Jetson line with WIS. I'm almost certain it will blow-back on the project and cause support headaches for us while all the while providing a sub-optimal user experience.

[barkingcat]

I have a Jetson as well, and you are sorely mistaken. Just reading the doc pages everything seems nice and well, but Nvidia deprecates these little boards like no other. No support after you've bought the thing, and everything is kept frozen. (ie no new python, no new python dependencies, etc) What they aren't telling you is that specific sub-versions within each jetson/orin family board have differing support (ie not what they say on that website you are reading), and it's up to you to figure it out.

I've gotten my Jetson to work well using Yocto to build my own linux distros with correct updated dependencies, libraries and updated jetpack, but it's not for the faint of heart, and that's a whole other ball of yarn. It also takes a few hours to generate a new build every time I need to update some dependency that depends on other dependencies (Yocto maintenance is a full time job in many embedded development shops - you're basically authoring your own distribution).

Treat these devices as what they are: embedded target boards for fixed industrial development (for example, to go into a robot or a car - once that design is finished, Nvidia will expect you to NEVER update any part of the system with an embedded jetson or orin system for years, until you replace the whole thing with their newest model that you buy off the shelf again).

This is standard fare in embedded and robotics space. Do not use these boards for any kind of rapidly moving software development, because it's the wrong tool for the job.

[kkielhofner]

+1

Software for Jetson boards should be viewed as firmware for these embedded/industrial devices. They get installed in a robot, MRI machine, etc with a specific bespoke application targeting what they came with and are never touched again -or- supported by some large commercial firm with the skillsets you describe.

I was as firm/absolute in my original reply as I was because anyone who thinks life with Jetson is similar to life with a discrete Nvidia GPU on x86_64 will be in for a huge shock and 95% of the time it will end up on their shelf in a year or two.

It's one thing when it's the latest random ARM SBC you bought for $50 with no vendor support, it's another thing entirely when you're spending > $600 (or $2000 as this started!!!!) on a Jetson.

[KeplerBoy]

Yes, it's all rather recent in my experience. You get CUDA 12 and the newest Pytorch.

[kkielhofner]

For now. Check back in a couple of years.