Styjun

Commencez à vous connecter -- I don't understand the phrasing of this

What's the reason for the decade jump in the recent X-Men trilogy?

Why did the AvroCar fail to fly above 3 feet?

Can Mage Hand be used to indirectly trigger an attack?

What is the theme of analysis?

ISP is not hashing the password I log in with online. Should I take any action?

Can Dive Down protect a creature against Pacifism?

Is it possible to install Firefox on Ubuntu with no desktop enviroment?

Why does this Apple //e drops into system monitor when booting?

What is the color associated with lukewarm?

What do you call the action of "describing events as they happen" like sports anchors do?

Harley Davidson clattering noise from engine, backfire and failure to start

Idiom for 'person who gets violent when drunk"

Dedicated bike GPS computer over smartphone

I sent an angry e-mail to my interviewers about a conflict at my home institution. Could this affect my application?

Why did the Death Eaters wait to reopen the Chamber of Secrets?

What publication claimed that Michael Jackson died in a nuclear holocaust?

Opposite of "Concerto Grosso"?

Is fission/fusion to iron the most efficient way to convert mass to energy?

I received a gift from my sister who just got back from

Parallelized for loop in Bash

Are athletes' college degrees discounted by employers and graduate school admissions?

Must a CPU have a GPU if the motherboard provides a display port (when there isn't any separate video card)?

Is it true that "only photographers care about noise"?

How to do large matrix decomposition with GPU in Tensorflow

TensorFlow: How to measure how much GPU memory each tensor takes?In tensorflow, how does one access a scalar tensor value before it is moved to the GPU?DMA between CPU and GPU in TensorFlowCPU/GPU Memory Usage with TensorflowTensorFlow GPU memoryHow does TensorFlow use both shared and dedicated GPU memory on the GPU on Windows 10?tensorflow not using gpu - prime number programtf.device('CPU:0') still utilizes gpu memorywhy Tensorflow-gpu is still using cpu

.everyoneloves__top-leaderboard:empty,.everyoneloves__mid-leaderboard:empty,.everyoneloves__bot-mid-leaderboard:empty height:90px;width:728px;box-sizing:border-box;

1

I am trying to do a matrix decomposition (or tucker decomposition on a tensor) in Tensorflow with GPU. I have tensorflow-gpu, my NVidia GPU has 4GB RAM. My problem is that my input matrix is huge, millions of rows and millions of columns and the size of the matrix is more than 5GB in memory. So each time Tensorflow gives me an out of memory (OOM) error. (If I turn off GPU, the whole process can run successfully in CPU using system RAM. Of course, the speed is slow.)

I did some research on Tensorflow and on NVidia CUDA lib. CUDA seems has a "unified memory" mechanism so the system RAM and GPU RAM share one address book. Yet no further details found.

I wonder if Tensorflow supports some memory sharing mechanism such that I can generate input in system RAM? (Since I want to use GPU to accelerate the calculations) And GPU can do the calculation piece by piece.

tensorflow gpu

share|improve this question

edited Mar 25 at 16:31

Miroslav Glamuzina

2,88521223

asked Mar 25 at 1:26

AmartinAmartin

62

• 
  
  
  

  
  

  
  
  
  
  
  

  
  
  Can this decomposition be split into smaller fragments manually? That’s your best bet. Shared memory architectures usually starve the GPU and you lose a lot of the speed advantage.
  
  – Kuba Ober
  Mar 25 at 16:36
  

  
  
  
  


• 
  
  
  

  
  

  
  
  
  
  
  

  
  
  Thank you for your cut-to-the-point comments, Kuba. I haven't found any easy/efficient ways to do it. One reason is I didn't find an easy way to split this SVD-similar job. Secondly, I am not sure it's worth it considering the difference of different RAM's bandwidth.
  
  – Amartin
  Mar 26 at 16:45
  

  
  
  
  

add a comment | 

1

I am trying to do a matrix decomposition (or tucker decomposition on a tensor) in Tensorflow with GPU. I have tensorflow-gpu, my NVidia GPU has 4GB RAM. My problem is that my input matrix is huge, millions of rows and millions of columns and the size of the matrix is more than 5GB in memory. So each time Tensorflow gives me an out of memory (OOM) error. (If I turn off GPU, the whole process can run successfully in CPU using system RAM. Of course, the speed is slow.)

I did some research on Tensorflow and on NVidia CUDA lib. CUDA seems has a "unified memory" mechanism so the system RAM and GPU RAM share one address book. Yet no further details found.

I wonder if Tensorflow supports some memory sharing mechanism such that I can generate input in system RAM? (Since I want to use GPU to accelerate the calculations) And GPU can do the calculation piece by piece.

tensorflow gpu

share|improve this question

edited Mar 25 at 16:31

Miroslav Glamuzina

2,88521223

asked Mar 25 at 1:26

AmartinAmartin

62

• 
  
  
  

  
  

  
  
  
  
  
  

  
  
  Can this decomposition be split into smaller fragments manually? That’s your best bet. Shared memory architectures usually starve the GPU and you lose a lot of the speed advantage.
  
  – Kuba Ober
  Mar 25 at 16:36
  

  
  
  
  


• 
  
  
  

  
  

  
  
  
  
  
  

  
  
  Thank you for your cut-to-the-point comments, Kuba. I haven't found any easy/efficient ways to do it. One reason is I didn't find an easy way to split this SVD-similar job. Secondly, I am not sure it's worth it considering the difference of different RAM's bandwidth.
  
  – Amartin
  Mar 26 at 16:45
  

  
  
  
  

add a comment | 

I am trying to do a matrix decomposition (or tucker decomposition on a tensor) in Tensorflow with GPU. I have tensorflow-gpu, my NVidia GPU has 4GB RAM. My problem is that my input matrix is huge, millions of rows and millions of columns and the size of the matrix is more than 5GB in memory. So each time Tensorflow gives me an out of memory (OOM) error. (If I turn off GPU, the whole process can run successfully in CPU using system RAM. Of course, the speed is slow.)

I did some research on Tensorflow and on NVidia CUDA lib. CUDA seems has a "unified memory" mechanism so the system RAM and GPU RAM share one address book. Yet no further details found.

I wonder if Tensorflow supports some memory sharing mechanism such that I can generate input in system RAM? (Since I want to use GPU to accelerate the calculations) And GPU can do the calculation piece by piece.

tensorflow gpu

share|improve this question

edited Mar 25 at 16:31

Miroslav Glamuzina

2,88521223

asked Mar 25 at 1:26

AmartinAmartin

62

share|improve this question

edited Mar 25 at 16:31

Miroslav Glamuzina

2,88521223

asked Mar 25 at 1:26

AmartinAmartin

62

share|improve this question

edited Mar 25 at 16:31

Miroslav Glamuzina

2,88521223

asked Mar 25 at 1:26

AmartinAmartin

62

edited Mar 25 at 16:31

Miroslav Glamuzina

2,88521223

edited Mar 25 at 16:31

Miroslav Glamuzina

2,88521223

asked Mar 25 at 1:26

AmartinAmartin

62

asked Mar 25 at 1:26

AmartinAmartin

62