💾 Archived View for dioskouroi.xyz › thread › 24919565 captured on 2020-10-31 at 00:53:08. Gemini links have been rewritten to link to archived content
-=-=-=-=-=-=-
________________________________________________________________________________
I have an NVIDIA Jetson Nano 4GB, bought it for ~$99 on Amazon[1] last December. The docs website has tutorials on how to install TensorFlow and it's pretty easy to get started. I got it to start playing with TF and it certainly did the job but I found that I was running into its limitations pretty quickly.
I (foolishly?) hoped I could run a low-param version of GPT-2, but even the smaller models were too large to run on this small GPU. So be aware if you're interested in this developer kit that it won't be able to run complex models, especially if you're getting the new 2GB version. Anything like image recognition/classification would work just fine though.
[1]
https://www.amazon.com/dp/B084DSDDLT
They have more powerful versions if you are willing to pay more than the $60
https://developer.nvidia.com/embedded/jetson-agx-xavier-deve...
I think this is a more appropriate step up
https://developer.nvidia.com/embedded/jetson-xavier-nx-devki...
It's a half Nintendo Switch without input and outputs. Half the GPU cores, half the RAM, same CPU.
recent discussion:
https://news.ycombinator.com/item?id=24687572
Haven't they had a 4GB kit out for a while? Why is everyone excited about a 2GB kit?
The price. $60 for that power at that form factor is really interesting.
4GB for $60:
https://sfbay.craigslist.org/pen/sys/d/san-francisco-nvidia-...
That doesn't help if you have to order 1000 of them.
The dev kit is just that - a dev kit. For production you could use the jetson nano module but probably want a different carrier board. I think the Nvidia license also prevents production use of the dev kit.
Seal up your electronics box and they'll never know. If it is not observable, it doesn't exist.
The alternative carrier boards all cost far too much, many more than the Nano itself which violates the first law of carrier boards [0]. Only the OEM carrier board is priced reasonably.
[0] A grocery bag should never cost more than the things that fit in it; a phone case should never cost more than its phone; a camera bag should never cost more than a camera; likewise, a carrier board should never cost more than what it carries.
You must not be familiar with the luxury handbag market.
Correct, I don't touch luxury bags for the same reason I don't touch Connect-Tech and Leopard Imaging products. I pay for functionality and carrier boards are functionally just wrappers with some wires.
There's no NV license on this. Note however that the carrier board is non-production specification and its warranty applies to development use only.
You don’t want to buy via distributor as a non-hobby person anyway. NVIDIA have a strong partner programs that hook you up with boards you need and with different pricing.
How much do those partner programs save you though?
For 1000 boards, from a distributor you're paying 1000 * $99 = $99000 and you are done with the order in 2 minutes.
To go through the partner program, hiring someone to have coffee and meetings and phone calls with NVIDIA for 6 months = $50000
Will going through the partner programs save you more than $50000?
Or can NVIDIA provide a pushbutton NLP bot that navigates their partner program for you and deals with all the necessary phone calls such that 1000 Nanos arrive at your doorstep for less than $99000?
It is more like email or two to get what you need. No need to employ fulltime person to handle it.
I just ordered from Arrow to Sweden. 600 SEK incl. VAT and shipping from US (DHL and FedEx are $0 for some reason)?!
This thing is going to crush it for my MMO, with the raspberry 4 I can render 100 characters at 60 FPS, I'm really curious what this thing will perform like!
I still don't see the value, unless I can train production size models it's not worth it. For IOT applications using CV etc, embedded AI in Arduino and Raspberry Pi is good enough.
I worked on optical sorting project that use cameras to scan many 1000s of small berries a second and you had around 10ms to extract each berry from video stream, calculate the center of object, decide if quality is good enough and schedule air stream to shoot in particular time to so it gets shoot out in exact time. Good luck trying to do that on Arduino type of board.
Industrial and industrial IoT is one of the use case for these kind of boards. These machines are not produced in high quantities so finding existing modules saves a tons of R&D budget.
Any camera modules you might recommend for such tasks, ideally compatible with the nano?
I can't remember the name of vendor, but they were from some niche manufacturer of professional machine vision cameras from Germany. Each of those cost few 1000s euros.
Almost anything for industrial use cost you a leg and then some. If you want to experiment with CV at home I would take a look at that high quality camera module from RPi.
Is the HQ camera module any better than the Veye Starvis (imx297/307/327) module?
Aside from being a third party module, they do quite well with low light situations -compared to conventional Raspberry Pi camera modules.
Blueberries?
Some applications require high speed inference, to be able to run segmentation and classification at high framerates
Not to mention doing other accelerated pre-processing & processing in addition to inference at the edge. There's certainly a use case for the Nano over Arduino, RPi, etc.
If what you are using is good enough than this might not interest you. Isn't that obvious?
Does this mean they are shipping? Still states "ship when ready" on pre-order status. Pretty excited to experiment. What peripherals are essential for CV, robotics, etc?
Mine shipped today, think 10/30 is the official release date
edit: mine too. due nov 3 election day ;)
I am surprised to see that there is one in stock at my local Micro Center.
Why was this thread even posted? Shameless nvidia advertising. Today belongs to AMD.