Affordable AI Kit for Developers and Hobby

Emperor

Juggernaut
Nvidia has launched the Jetson Orin Nano Super Developer Kit, a compact generative AI supercomputer designed for developers, hobbyists, and students. Priced at USD 249 (down from USD 499), the new kit delivers a 1.7x increase in generative AI inference performance and a 70 percent boost to 67 INT8 TOPS. Additionally, it provides a 50 percent increase in memory bandwidth, now reaching 102 GBps compared to its predecessor, Nvidia said on Tuesday. More details
 
Few years back, I used to toy around with RPi 3B+ and most of the efforts went into optimising the models for RPi. Nvidia’s nano kit definitely is a nice alternative if full fledged system is not an option.
 
The original Jetson Orin nano is still being sold for 50k or 580 USD


So to get a more powerful version at half the cost in stock would be a quite difficult. If it becomes available at say 25-30k I would be quite interested in getting one. Would be great for running a local LLM with a power consumption below 20 watts.
I believe at that time it was available for around 12k
 
I believe at that time it was available for around 12k

Jetson platform has a lot of models available. There a huge difference in power among the models. I don't believe a SBC model of Jetson Orin Nano was ever available for 12k. Lowest was a $199 dollar version for a module which needs a carrier board which will cost you more than a mid ranged AM5 motherboard.

In India this 4gb module introduced in 2023 is 25k and will need another carrier board like a Avermedia D133 which will cost another 14k.


If you can get a more powerful newer version SBC for 249 USD. That a good deal.

Looking at specs the performance difference is double
Orin Nano had 512 cuda core and 16 tensor cores.
Orin Nano Super has 1024 cuda cores and 32 tensor cores.
Orin AGX has 2024 cuda core and 64 tensor cores. (cost is usually over 1 lakh)
RTX 4060 GPU has 3072 cuda cores.

For older version of Jetson nano you can get it for around 5k for the lowest 2gb model. This model with 128 core maxwell GPU and a quad core CPU would be more powerful than a Raspberry Pi 4 for GPU intensive applications like AI, machine learning task.

In terms of AI
Jetson Nano Super does 68 INT8 TOPS
Intel ARC B580 does 233 INT8 TOPS (same price as jetson)
Nvidia 4060 does 242 INT8 TOPS

INT8 TOPS is a measure of 8 bit integer operations per second measured in trillions. So 68 trillion 8 bit integer operations executed per second for this SBC which consumes less than 20 watts of power.