Machine Learning on RX VEGA

[GirlFromYonder]

I will be choosing AI in university and i was wondering that for Machine learning proposes can I use a consumer grade AMD gpu and not the Instinct ones cause i want to save money and am a poor cheapskate University student. I mean 4 RX Vegas compared to one Instinct card would be better?

[Hunter-97-G]

Yeah, probably, provided Crossfire scales well with deep learning.

[thewipyk]

Might as well buy a used nvidia Tesla accelerator card as they are designed for this. They aren't graphics card, rather accelerators. However, AMD vega is also good choice as they advertise it as machine learning card

[Devin92]

45 minutes ago, thewipyk said:

Might as well buy a used nvidia Tesla accelerator card as they are designed for this. They aren't graphics card, rather accelerators. However, AMD vega is also good choice as they advertise it as machine learning card

is AMD GPU widely supported by popular deep learning library now? I know Theano still doesnt and tensorflow probly does not either.

[thewipyk]

1 hour ago, Devin92 said:

is AMD GPU widely supported by popular deep learning library now? I know Theano still doesnt and tensorflow probly does not either.

I'm not really into deep learning stuff, sorry.

[pyrojoe34]

2 hours ago, Hunter-97-G said:

Yeah, probably, provided Crossfire scales well with deep learning.

Currently, computational work usually does not use things like crossfire and SLI, those are used for basically exclusively for gaming. They typically work on split tasks that, once completed, are usually consolidated by the CPU. Scaling can still be an issue but properly designed computational tasks can scale at near 100% given that individual processes are typically not shared across GPUs.

[Devin92]

27 minutes ago, pyrojoe34 said:

Currently, computational work usually does not use things like crossfire and SLI, those are used for basically exclusively for gaming. They typically work on split tasks that, once completed, are usually consolidated by the CPU. Scaling can still be an issue but properly designed computational tasks can scale at near 100% given that individual processes are typically not shared across GPUs.

I know for some CNN (convolutional neural network) training, multi-GPU is introduced almost a decade ago, where each GPU is responsible for training a certain branch of the network and the scaling is pretty good. CF and SLI is not quiet useful here since those are designed to finish one task utilizing multi-GPUs, where deep learning have multiple tasks in training and it can be very well separated to multi-GPU.

 

The only drawback for AMD is the support of libraries. Developers are adopting AMD GPUs but to be on the safe side, just try to find good deal on Nvdia's GPU which have a good amount of CUDA cores cause this is what accelerate deep learning training process.

 

[pyrojoe34]

9 minutes ago, Devin92 said:

I know for some CNN (convolutional neural network) training, multi-GPU is introduced almost a decade ago, where each GPU is responsible for training a certain branch of the network and the scaling is pretty good. CF and SLI is not quiet useful here since those are designed to finish one task utilizing multi-GPUs, where deep learning have multiple tasks in training and it can be very well separated to multi-GPU.

 

The only drawback for AMD is the support of libraries. Developers are adopting AMD GPUs but to be on the safe side, just try to find good deal on Nvdia's GPU which have a good amount of CUDA cores cause this is what accelerate deep learning training process.

 

Yea support is a huge issue. I know our computational core here uses a molecular docking program that only supports CUDA so they have no choice but to go NVIDIA (we have a couple racks with 2x12 cores Xeons and 8 GTX1080s which I'm pretty jealous of). It seems like openCL should become more universally utilized so that these programs are not limited to CUDA.

[Devin92]

1 hour ago, pyrojoe34 said:

Yea support is a huge issue. I know our computational core here uses a molecular docking program that only supports CUDA so they have no choice but to go NVIDIA (we have a couple racks with 2x12 cores Xeons and 8 GTX1080s which I'm pretty jealous of). It seems like openCL should become more universally utilized so that these programs are not limited to CUDA.

24 CPU cores, and 8 1080? I am not just pretty jealous...... I am furiously jealous..... here I am useing i7 and a gtx1060.

[GirlFromYonder]

Ok so 2-4 gtx 1080 ti will be better than 2-4 vega cards?

[GirlFromYonder]

13 hours ago, Devin92 said:

I know for some CNN (convolutional neural network) training, multi-GPU is introduced almost a decade ago, where each GPU is responsible for training a certain branch of the network and the scaling is pretty good. CF and SLI is not quiet useful here since those are designed to finish one task utilizing multi-GPUs, where deep learning have multiple tasks in training and it can be very well separated to multi-GPU.

 

The only drawback for AMD is the support of libraries. Developers are adopting AMD GPUs but to be on the safe side, just try to find good deal on Nvdia's GPU which have a good amount of CUDA cores cause this is what accelerate deep learning training process.

i most probably will be build somewhere around december or next year. And knowing the industrial. I dont think it will adopt OpenCL anytime soon.

 

 

So Nvidia it is?

[Devin92]

6 hours ago, DeezNoNos said:

i most probably will be build somewhere around december or next year. And knowing the industrial. I dont think it will adopt OpenCL anytime soon.

 

 

So Nvidia it is?

yeap. I have seen Theano start adopting OpenCL for almost half a year, and their support is still minimal.

[GirlFromYonder]

2 hours ago, Devin92 said:

yeap. I have seen Theano start adopting OpenCL for almost half a year, and their support is still minimal.

Shitr well lol. i wanted to buy a freesync monitor with that. well i guess no frame sync for my games then.