Tensorflow

[Bigun]

So, I'm in the market for a GPU specifically for machine learning, and it's going into a headless server, so I really do not care about 3D gaming as far as this rig goes.

I need at least CUDA 3 compatibility (obviously the higher the better), the higher memory and processor power also helps.  But I'm unsure of which one to get, and what kind of spending I could expect.

Honestly I think my limit is about $250 USD, but if I need to go higher, I can wait a little bit and save what I need.  Any help is appreciated.

[Electric_Boi]

34 minutes ago, Bigun said:

-snip-

Save up a bit and go with GTX 1060 6 GB

[Radium_Angel]

42 minutes ago, Bigun said:

So, I'm in the market for a GPU specifically for machine learning, and it's going into a headless server, so I really do not care about 3D gaming as far as this rig goes.

I need at least CUDA 3 compatibility (obviously the higher the better), the higher memory and processor power also helps.  But I'm unsure of which one to get, and what kind of spending I could expect.

Honestly I think my limit is about $250 USD, but if I need to go higher, I can wait a little bit and save what I need.  Any help is appreciated.

I do machine learning as a semi-serious hobby, and unless you are trying to make a living at this, any GFX card beyond the 660 GTX will do you fine. I've tested many cards, from old Radeon 4870s to modern ones, and the differences are slight. 

Now, a disclaimer, I work in the textural field (ie, not image or sound recognition) what kind of ML are you interested in?

[Bigun]

2 minutes ago, Radium_Angel said:

I do machine learning as a semi-serious hobby, and unless you are trying to make a living at this, any GFX card beyond the 660 GTX will do you fine. I've tested many cards, from old Radeon 4870s to modern ones, and the differences are slight. 

Now, a disclaimer, I work in the textural field (ie, not image or sound recognition) what kind of ML are you interested in?

Currently LSTM text generation based off of someone's Tweets.  But who knows in the future, I'm still learning.

[Radium_Angel]

8 minutes ago, Bigun said:

Currently LSTM text generation based off of someone's Tweets.  But who knows in the future, I'm still learning.

Then upgrade GFX cards when you are ready.

This site might help

http://aiweirdness.com/

 

She does her work on an old macbook. 

[Bigun]

10 minutes ago, Radium_Angel said:

Then upgrade GFX cards when you are ready.

This site might help

http://aiweirdness.com/

 

She does her work on an old macbook. 

That's the problem.  It's CPU only as the machine doesn't have an Nvidia graphics card.  One epoch is taking 3 hours.

[Radium_Angel]

2 minutes ago, Bigun said:

That's the problem.  It's CPU only as the machine doesn't have an Nvidia graphics card.  One epoch is taking 3 hours.

I am aware that being CPU-bound slows things down, but the point being that the GFX card, while making a big difference, the *spped of the GFX card* is less important.

[Bigun]

17 minutes ago, Radium_Angel said:

I am aware that being CPU-bound slows things down, but the point being that the GFX card, while making a big difference, the *spped of the GFX card* is less important.

I agree, so suggestions on make/model?  Perhaps Electric_Boi's suggestion of a GTX 1060 6 GB?

[Radium_Angel]

6 hours ago, Bigun said:

I agree, so suggestions on make/model?  Perhaps Electric_Boi's suggestion of a GTX 1060 6 GB?

Overkill, unless you want to spend the cash. I do this:

On a GTX 660

 

(it's music generated from lyrics, which in turn has been created by my RNN. It's no great shakes, but some of the tunes are pretty cool, Check out "Communist Manifesto")

 

[Bigun]

Ok, I've made the jump, GTX 660 on the way.

[Radium_Angel]

1 hour ago, Bigun said:

Ok, I've made the jump, GTX 660 on the way.

Cool, happy to help with your journey if you need it, feel free to PM me if you need it

[Bigun]

12 hours ago, Radium_Angel said:

Cool, happy to help with your journey if you need it, feel free to PM me if you need it

With machine learning:  always.  When I get to a point where I feel I've hit a brick wall, I'll be contacting you.

[Radium_Angel]

32 minutes ago, Bigun said:

With machine learning:  always.  When I get to a point where I feel I've hit a brick wall, I'll be contacting you.

Cool. Look into Char-RNN by karpathy for a good start on RNN software. You'll need Torch and Torch-CL, but they aren't hard to install.

 

[Bigun]

20 minutes ago, Radium_Angel said:

Cool. Look into Char-RNN by karpathy for a good start on RNN software. You'll need Torch and Torch-CL, but they aren't hard to install.

 

Right now I'm using tensorflow/keras.  After I've mastered that, I may move on to torch.

[Bigun]

So, while waiting for the new card to arrive, I went to work this morning, and was able to scrape up a GeForce GT 710 w/ 1 GB DDR3, it's a bit faster than CPU, but only a little.  One Epoch with my scripting on a 4-core was taking an hour, this is taking about 35 minutes.

 

It looks like it's using the GPU:

 

2018-07-09 12:32:40.152653: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:898] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2018-07-09 12:32:40.153162: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1356] Found device 0 with properties:
name: GeForce GT 710 major: 3 minor: 5 memoryClockRate(GHz): 0.954
pciBusID: 0000:01:00.0
totalMemory: 980.00MiB freeMemory: 661.56MiB
2018-07-09 12:32:40.153202: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1435] Adding visible gpu devices: 0
2018-07-09 12:32:40.381046: I tensorflow/core/common_runtime/gpu/gpu_device.cc:923] Device interconnect StreamExecutor with strength 1 edge matrix:
2018-07-09 12:32:40.381092: I tensorflow/core/common_runtime/gpu/gpu_device.cc:929]      0
2018-07-09 12:32:40.381100: I tensorflow/core/common_runtime/gpu/gpu_device.cc:942] 0:   N
2018-07-09 12:32:40.381265: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1053] Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:0 with 398 MB memory) -> physical GPU (device: 0, name: GeForce GT 710, pci bus id: 0000:01:00.0, compute capability: 3.5)

But the performance is so poor I wanted to make sure.

[Bigun]

So the card arrived, and I got bit:

 

 $ python3.6 list_devices.py
2018-07-14 11:10:51.082115: I tensorflow/core/platform/cpu_feature_guard.cc:141] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.1 SSE4.2
2018-07-14 11:10:51.372493: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:897] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2018-07-14 11:10:51.373123: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1404] Found device 0 with properties:
name: GeForce GTX 660 major: 3 minor: 0 memoryClockRate(GHz): 1.137
pciBusID: 0000:01:00.0
totalMemory: 2.95GiB freeMemory: 2.91GiB
2018-07-14 11:10:51.373156: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1454] Ignoring visible gpu device (device: 0, name: GeForce GTX 660, pci bus id: 0000:01:00.0, compute capability: 3.0) with Cuda compute capability 3.0. The minimum required Cuda capability is 3.5.
2018-07-14 11:10:51.373178: I tensorflow/core/common_runtime/gpu/gpu_device.cc:964] Device interconnect StreamExecutor with strength 1 edge matrix:
2018-07-14 11:10:51.373192: I tensorflow/core/common_runtime/gpu/gpu_device.cc:970]      0
2018-07-14 11:10:51.373206: I tensorflow/core/common_runtime/gpu/gpu_device.cc:983] 0:   N
[name: "/device:CPU:0"
device_type: "CPU"
memory_limit: 268435456
locality {
}
incarnation: 7428093072061845487
]

Looks like another card or degrade tensorflow.