Open HanSeYeong opened 5 years ago
If I connect two TITAN RTX with nvlink, will my ram double? and can I use that ram in training?
No (as I know)
My dataset resolution is 1280X720 so I can choose either increasing the width and height of network or decreasing subdivisions. I wanna recognize special trucks which is reasonable big. In this situation, which option do you recommend to increase accuracy?
If your objects are big, then you shouldn't set high network resulotion.
Just train two models with different cfg-files and check where do you get the highest mAP (accuracy).
@AlexeyAB No double ram?
Then I doubt how could you make your trained weight with width and height are 608 and subdivisions=16? In my experience 23GB of memory will be required with that configuration.
I wanna know what graphic card did you use before. Thanks to read my question.
@AlexeyAB They are advertising that nvlink can make double ram, so I have no doubt that I can't use 48GB of memory to train YOLO model....
@HanSeYoung
Then I doubt how could you make your trained weight with width and height are 608 and subdivisions=16? In my experience 23GB of memory will be required with that configuration.
If you want to have 24 GB GPU-RAM, then you can use:
They are advertising that nvlink can make double ram, so I have no doubt that I can't use 48GB of memory to train YOLO model....
@AlexeyAB They are advertising that nvlink can make double ram, so I have no doubt that I can't use 48GB of memory to train YOLO model....
I have never tried this.
I do not know whether it is possible to force the CUDA and cuDNN library to see two cards as one or two cards with a single doubled memory.
And will not the interaction between the GPU greatly slow down performance, since nvLink (100 GB/sec), while GPU-RAM (~500 GB/sec)
If you can combine the memory of two RTX cards and successfully launch Darknet with lower subdivisions
, and measure the acceleration or deceleration performance, then let me know.
Thanks @AlexeyAB
I think you are right. The speed bottleneck will be matter even if I can make them be configured by one GPU.
But I'll buy one more titan to accelerate training speed and try if I can use double ram.
Really appreciate.
@HanSeYoung
It seems Turing GPUs can share their VRAM.
https://www.techpowerup.com/reviews/NVIDIA/GeForce_RTX_2080_Ti_SLI_RTX_2080_SLI_NVLink/9.html
With NVLink things have changed dramatically. All SLI member cards can now share their memory with the VRAM of each card sitting at different addresses in a flat address space. Each GPU is able to access the other's memory
So if there is bottleneck in the VRAM-amount rather than VRAM-bandwidth, then it can improve performance, for example, by using larger mini-batch that is speedup training.
@AlexeyAB Yeah that's what I want from NVLINK!
But as I searched, I couldn't find how to use double vrams by NVLINK.... If you find the solution please share to all of YOLO lovers!
Thanks to notify me!
@HanSeYeong not sure if you have progressed on this or tried? But it seems RTX cards within Linux environment will support memory pooling with CUDA. Not sure how / if it will work with darknet, but curious to find out.
See below:
https://www.daz3d.com/forums/discussion/353011/test-if-nvlink-is-working-for-vram-pooling
I wonder if I can use 48GB of ram with 2 Titan RTX.
I trained my data with below configurations.
I also added some photos with blank text file and I got 92% accuracy which is really awesome.
But further question,
I am always grateful to you. Thanks