NVIDIA Jetson TK1 Host PC
Hello I plan on using the NIVDIA Jetson TK1 for testing and enhancing deep learning. But I want to ask which computers would be best in doing these type of procedures, I've read that the NVIDIA Jetson TK1 needs a host PC which integrated NVIDA graphic cards to enhance deep learning and machine learning. Thanks
Hello

I plan on using the NIVDIA Jetson TK1 for testing and enhancing deep learning. But I want to ask which computers would be best in doing these type of procedures, I've read that the NVIDIA Jetson TK1 needs a host PC which integrated NVIDA graphic cards to enhance deep learning and machine learning.

Thanks

#1
Posted 02/11/2018 07:18 AM   
A Linux host PC (x86_64) is required for flashing a Jetson. The same software which is used for flash can also install various programs and software on both the host and Jetson. This includes CUDA, development tools, so on. Only a computer with an NVIDIA GPU can be used with CUDA, so if you want those tools on the host (think of developing on the host and deploying on the Jetson), then you need that GPU. If you just flash or install to the Jetson, then you don't need the NVIDIA GPU on the host. Typically a Jetson is good at using pre-trained models, but you need something beefier for training. If your desktop has thousands of CUDA cores, and the Jetson has only 256, you'll see a much better training on the larger GPU (and also typically the desktop GPU may have more video RAM...I'd go 6GB or more, such as a GeForce 1060 or faster...the Jetson uses only a subset of its installed system RAM without dedicated video/GPU RAM). The difference is that you might want a 1000 watt power supply on a fast workstation, while the Jetson is using 7.5 to 30W (depending on what is being done and which Jetson you are using). The TK1 has its most distinguishing limitation from being 32-bit. CUDA development won't go beyond version 6.5 on any 32-bit platform. Newer CUDA is constantly evolving on the 64-bit versions. So if you want to keep up with modern software, then you need a TX1 at minimum, or even better yet a TX2. If you need the compute power with less energy drain (e.g., a battery powered drone), then go with the TX2. It is astonishing how well this little electrical supply does so much...if you just need raw power, and if you are ok with drawing 10 to 30 times more power drain, then go with a desktop computer.
A Linux host PC (x86_64) is required for flashing a Jetson. The same software which is used for flash can also install various programs and software on both the host and Jetson. This includes CUDA, development tools, so on. Only a computer with an NVIDIA GPU can be used with CUDA, so if you want those tools on the host (think of developing on the host and deploying on the Jetson), then you need that GPU. If you just flash or install to the Jetson, then you don't need the NVIDIA GPU on the host.

Typically a Jetson is good at using pre-trained models, but you need something beefier for training. If your desktop has thousands of CUDA cores, and the Jetson has only 256, you'll see a much better training on the larger GPU (and also typically the desktop GPU may have more video RAM...I'd go 6GB or more, such as a GeForce 1060 or faster...the Jetson uses only a subset of its installed system RAM without dedicated video/GPU RAM). The difference is that you might want a 1000 watt power supply on a fast workstation, while the Jetson is using 7.5 to 30W (depending on what is being done and which Jetson you are using).

The TK1 has its most distinguishing limitation from being 32-bit. CUDA development won't go beyond version 6.5 on any 32-bit platform. Newer CUDA is constantly evolving on the 64-bit versions. So if you want to keep up with modern software, then you need a TX1 at minimum, or even better yet a TX2. If you need the compute power with less energy drain (e.g., a battery powered drone), then go with the TX2. It is astonishing how well this little electrical supply does so much...if you just need raw power, and if you are ok with drawing 10 to 30 times more power drain, then go with a desktop computer.

#2
Posted 02/11/2018 05:34 PM   
Most PCs that I've search on the net that have an NVIDIA GPU seem to be PCs intended for gaming. Is it reasonable to buy such a computer if I'm only using it for installing and running software applications such as CUDA, NVcaffe, DIGITS, and installing other Nvidia drivers? I'll need a host PC when working with the TK1 as it speeds the training process and I'd like to follow the instructions that tell me to do so.
Most PCs that I've search on the net that have an NVIDIA GPU seem to be PCs intended for gaming. Is it reasonable to buy such a computer if I'm only using it for installing and running software applications such as CUDA, NVcaffe, DIGITS, and installing other Nvidia drivers?

I'll need a host PC when working with the TK1 as it speeds the training process and I'd like to follow the instructions that tell me to do so.

#3
Posted 02/12/2018 01:21 AM   
Yes, the gaming PCs with NVIDIA graphics are probably for training so far as ordinary purposes are concerned. The 1080Ti would be a rather fast substitute versus training on a Jetson. More video RAM would be good, e.g., an 11GB 1080Ti would have a benefit versus say 6GB on something else even if the basic card behind the RAM was the same. Some of the deep learning can use more memory than what a game might use...and we definitely like games (it wouldn't hurt for that either). Once you go past a 1080Ti cost can go up fast. The 1060 would be a good buy if you are on a budget, but get the 6GB version...3GB won't do too well. Consider that gaming and deep learning have a lot in common...unless you do something more exotic a gaming PC is probably one of the better choices. Having this as your PC would allow training, and would also allow cross compile and testing (you would need to tell the compile to build for both GPUs, but it should be fairly simple...JetPack will install nsight on the host PC if you want and set up some of the cross-environment).
Answer Accepted by Forum Admin
Yes, the gaming PCs with NVIDIA graphics are probably for training so far as ordinary purposes are concerned. The 1080Ti would be a rather fast substitute versus training on a Jetson. More video RAM would be good, e.g., an 11GB 1080Ti would have a benefit versus say 6GB on something else even if the basic card behind the RAM was the same. Some of the deep learning can use more memory than what a game might use...and we definitely like games (it wouldn't hurt for that either). Once you go past a 1080Ti cost can go up fast. The 1060 would be a good buy if you are on a budget, but get the 6GB version...3GB won't do too well.

Consider that gaming and deep learning have a lot in common...unless you do something more exotic a gaming PC is probably one of the better choices.

Having this as your PC would allow training, and would also allow cross compile and testing (you would need to tell the compile to build for both GPUs, but it should be fairly simple...JetPack will install nsight on the host PC if you want and set up some of the cross-environment).

#4
Posted 02/12/2018 07:32 PM   
Scroll To Top

Add Reply