Processing Unit Build for a Self-Driving Car

[Imperialdramon]

Budget (including currency): 3000 USD (flexible up to 10000 USD)

Country: United States

Games, programs or workloads that it will be used for: running CARMA

Other details:

Hi,

 

We are working on transforming a 2017 Ford Fusion Hybrid (seen below) into a connected and (partially) autonomous car that runs CARMA for our research. As part of that, we want to install a processing unit in the trunk in the empty space to the right of the aluminum frame in the image below.

 (

 

Here is what we are looking for in the processing unit:

 

Dimensions: ~ 400 * 300 * 200 mm. We plan on installing a Mikrotik CRS328-24P-4S+RM network switch at the top of the right cell.

Performance: Min: 5 TFLOPS, Desired: 10 TFLOPS or more. Capable of processing data from up to 8 1080p, 12bit IMX390 camera sensors, up to two 64-beam lidar sensors, and up to 4 Continental short- and long-range radar sensors.

Max Power Consumption: Desired: 180W, Max: 400W. Each port of the power distribution system (which is connected to the car's battery) provides 180W. Right now it only provides power to the sensors, but ideally we want it to power the processing unit as well. If the desired performance can't be reached with under 180W, ideas are welcome for how to power the system.

CPU Cores(threads): Min: 8(16), Desired: 12(24) or more.

GPU VRAM: Min: 6GB, Desired: 8GB or more. Anything that matches the performance requirement above. I guess a Geforce 30 series has the best combo of performance/TDP, but I know it's hard to come by.

RAM: Min: 16 GB, Desired: 32 GB or more.

Storage: Min: 10TB, Desired: 40TB or more. A combination of fast M.2 NVMe and HDD.

Cooling: When the trunk is closed the heat from the unit will turn the trunk into a furnace, so ideas for cooling the unit are really welcome!

Vibration: Not sure if we should be worried about this but any tips are welcome.

 

I'd appreciate your help recommending parts and cooling/power supply strategies for building this system. I was personally thinking about a combo of Ryzen 9 and RTX 3060 (assuming I can find any) that are underclocked to fit the power budget, but I'm curious to see what the community recommends.

 

[Enderman]

https://www.nvidia.com/en-us/self-driving-cars/drive-platform/hardware/

[dizmo]

Someone trying to make their own self driving car, coming to a forum largely populated by youths for advice.
This is everything I fear about automated vehicles.

[Imperialdramon]

32 minutes ago, Enderman said:

https://www.nvidia.com/en-us/self-driving-cars/drive-platform/hardware/

I'm aware of that, but they are really expensive, like in the order of $50,000. Besides, most of the price is for the Nvidia software which we don't need.

28 minutes ago, dizmo said:

Someone trying to make their own self driving car, coming to a forum largely populated by youths for advice.
This is everything I fear about automated vehicles.

Then don't use them! I don't think anyone's forcing you to.

On a serious note though, if you had read the post you would have known that the car is being built for research (i.e. only operated in university parking lots) and that I am asking for recommendations for processing hardware. You would have also known that the software, which does the actual driving, is developed by the FHWA. Apart from that, you need special tools to even give the car commands to drive itself, and it's not cheap. That's why you don't see everyone and their mom building one in their back yard.

[Enderman]

15 minutes ago, Imperialdramon said:

I'm aware of that, but they are really expensive, like in the order of $50,000. Besides, most of the price is for the Nvidia software which we don't need.

 

Maybe the Jetson AGX Xavier then?

[Imperialdramon]

20 minutes ago, Enderman said:

Maybe the Jetson AGX Xavier then?

It's not powerful enough, and you can't log a lot of data on it. We are using one to capture image frames from cameras (because they use the GMSL2 protocol, you can see it upside down at the bottom right corner of the image) and do some image encoding, but the processing needed for other parts, especially for lidar data needed for object detection and localization, is beyond its capacity.