In the past week or so, I have been building and setting up a new computer for the golf cart. Now, this new machine is fully set up and ready to go. Next week, I will finish building a cargo box on the back of the golf cart where the computer will be placed. I am very excited about the next level of the golf cart’s hardware.
The NVIDIA Jetson is a decent computer, why switch to a better one when I had very limited resources? I asked myself this question frequently in the past month. The conclusion is simple, the Jetson can no longer support the intense workload that a self-driving car requires. For example, the Jetson cannot even run the popular tiny-YOLO object detection network without exhausting all of its graphics card memory. The reason behind this computing threshold is largly due to the lack of RAM on the Jetson. Yes, this computer doesn’t have RAM. In fact, the Jetson treats the video card memory as RAM. Even though the card has 6GB, I can only effectively use ~3GB.
Another drawback of using the Jetson is the lack of CPU power. Since NVIDIA wants to keep the cost down and focus on the graphics card, the CPU of the Jetson can’t efficiently handle heavy processing.
I realized that in order to make significant progress, I need to make some investment for a new computer. It’s worth it.
The new machine
I focused on two things when I built this new computer.
1. keep the cost down.
2. Make sure it’s robust.
Maximizing those two things is rather paradoxical. Eventually, I came to a very satisfying conclusion with a decent computer. I have to say that working on a quality machine is integral to not only good engineering but also preserving my sanity. 😎
Here are the parts. (I didn’t need to buy a graphics card because Deerfield Academy also purchased one for this project back in the winter. )
Other than the computer itself, I needed to find a powerful battery that can power this ~220W beast. This is the most cost-efficient option I found on Amazon. This portable generator can power up to 300W and has a 200Wh Lithium battery. Meaning that if my computer uses 200W, the battery can power it for an hour. 👍
After completing the build, this is what the computer looks like on the inside.
Frankly, I am not disappointed nor exhilarated about the performance of this machine. It’s pretty much exactly what I expected.
The GeekBench single core score is: ~4300
The GeekBench multicore score is ~8400
The score above is very similar to my 2017 MacBook Pro, which is satisfying.
I also tested tiny-YOLO (object detection) and semantic segmentation on this new machine. (By the way, the computer can’t handle full YOLO). YOLO’s framerate was around 11fps. Segmentation’s framerate is around ~11 with visualization. Combining them together, plus the steering predictor and other processes, the framerate was around ~6fps. This is a significant improvement! The NVIDIA Jetson couldn’t even load the tiny-YOLO model without crashing! However, I believe that if I had the GTX1070 or GTX1080, the framerate will be >10fps. 💸
I would like to say a tremendous thank you to my parents, who financially supported me with this hardware upgrade. Without their help, I would be stuck with the Jetson.