What exactly are the requriments for machine learning?

[Dinnerb0n3]

Hey guys so I wanted to gift my brother a new GPU. He says he needs a new GPU because AMD cards are not well suited for machine learning and he has an Rx480.

Which GPU shoulders I buy that would let him do machine learning without many hickups and won't tear a hole in my wallet.

DISCLAIMER:I have no idea about machine learning.

[Linker]

There are no requirements, any computer can do it, it just depends on how fast it can do it.

All a new gpu would do is reduce the computation time.

I'd ask him (if the gift isn't a surprise.) how into it he is. If he is already having issues like needing to run calculations for 5+ hours, then it might benefit him to have a faster gpu.

Otherwise, I can't imagine an rx480 being awful at it unless he is doing huge workloads.

Imagine it like gaming- some games are hard to run and require top of the line hardware, some don't, and will run on a rx480 just fine.

 

DISCLAIMER: I haven't done machine learning work, like ever.

[Fasauceome]

Machine learning can be done on just about anything, but the hardware that does it well is gonna be something with tensor cores, anything RTX at the moment, or a Titan V if you win a lottery.

[Princess Luna]

8 minutes ago, fasauceome said:

Titan V if you win a lottery.

Exactly, The V is still the king of Machine Learning, Tensor Cores are the best thing ever for all software around AI and the V has more of those than even the RTX 2080 Ti (100 more).

 

RT Cores are useless, CUDA cores are 'fine'.

 

So yes, the cheapest yet greatest option for OP is the RTX 2060 and the top of the line option the TITAN V.

[Dinnerb0n3]

17 hours ago, Princess Cadence said:

Exactly, The V is still the king of Machine Learning, Tensor Cores are the best thing ever for all software around AI and the V has more of those than even the RTX 2080 Ti (100 more).

 

RT Cores are useless, CUDA cores are 'fine'.

 

So yes, the cheapest yet greatest option for OP is the RTX 2060 and the top of the line option the TITAN V.

How about the 2070 it is on sale at 465USD right now.

[Princess Luna]

1 minute ago, Dinnerb0n3 said:

How about the 2070 it is on sale at 465USD right now.

It will be a considerable better experience than a RX480

[Dinnerb0n3]

Just now, Princess Cadence said:

It will be a considerable better experience than a RX480

Ok thanks.

 

[Dinnerb0n3]

18 hours ago, Linker said:

There are no requirements, any computer can do it, it just depends on how fast it can do it.

All a new gpu would do is reduce the computation time.

I'd ask him (if the gift isn't a surprise.) how into it he is. If he is already having issues like needing to run calculations for 5+ hours, then it might benefit him to have a faster gpu.

Otherwise, I can't imagine an rx480 being awful at it unless he is doing huge workloads.

Imagine it like gaming- some games are hard to run and require top of the line hardware, some don't, and will run on a rx480 just fine.

 

DISCLAIMER: I haven't done machine learning work, like ever.

I believe that he said he wanted to simulate liquid flow or something along those lines.

[Gmoneyinthebank]

Nvidia just released the nano developer board for $99.  Comes with a Navier Stokes 3D incompressable fluid simulation package in their Jetpack.

https://developer.nvidia.com/embedded/buy/jetson-nano-devkit

 

That being said... Pretty much any computer can be used for ML.  Ex: Tensorflow has a CPU only version that is free

[Dinnerb0n3]

On 4/9/2019 at 5:28 PM, Gmoneyinthebank said:

Nvidia just released the nano developer board for $99.  Comes with a Navier Stokes 3D incompressable fluid simulation package in their Jetpack.

https://developer.nvidia.com/embedded/buy/jetson-nano-devkit

 

That being said... Pretty much any computer can be used for ML.  Ex: Tensorflow has a CPU only version that is free

Will the 2060 run it well? And how much of a performance gain will I get if I upgrade to the 2070.

Will the 6 GB vram be a bottleneck.

 

[BuckGup]

It’s basically all done in CUDA so AMD is out of the question unless you use their translation api. But it’s C/C++ with some extra bits added. Turing is the fastest for machine learning and you can utilize the tensor cores for some tasks. Otherwise any 2000 or 1000 series card works. 2000 will be faster and the more you spend the faster the training is.

[Gmoneyinthebank]

5 hours ago, Dinnerb0n3 said:

Will the 2060 run it well? And how much of a performance gain will I get if I upgrade to the 2070.

Will the 6 GB vram be a bottleneck.

 

It's a potential bottleneck but it's high enough that it's unlikely to ever matter to an independent programmer learning ML.

 

My advice is to get the 2060 and take the extra $100 to get an Nvidia Nano.  He's better off getting started ML/AI programming on a developer board and buying dedicated gaming equipment.

[Dinnerb0n3]

19 hours ago, Gmoneyinthebank said:

It's a potential bottleneck but it's high enough that it's unlikely to ever matter to an independent programmer learning ML.

 

My advice is to get the 2060 and take the extra $100 to get an Nvidia Nano.  He's better off getting started ML/AI programming on a developer board and buying dedicated gaming equipment.

Hmmm that sounds nice, but is there any other option other than the Jetson nano in the same price range.

Idk anything about ml or ai

[__orange]

On 4/9/2019 at 10:28 PM, Gmoneyinthebank said:

That being said... Pretty much any computer can be used for ML. 

Jetson Nano is designed for edge inferencing. It is not needed for deep learning (it is only for deploying). You still need a CUDA device for training. Just buy a nvidia (gaming) graphics card which fit your budget. Titan V/Titan RTX is fine but too expensive. Better to choose 2x RTX2080Ti. GTX graphics cards (without tensor cores) are also good if you can't afford a RTX2060.

 

On 4/12/2019 at 3:15 AM, Dinnerb0n3 said:

Hmmm that sounds nice, but is there any other option other than the Jetson nano in the same price range.

 Idk anything about ml or ai

For edge inferencing device like Jetson Nano, there are:

Raspberry Pi, Kirin 970/980, Hisilicon Hi3559a, NextVPU N171, Rockchip RK3399Pro/RK1808, Snapdragon 845/850, etc...

Do NOT buy if your are not familiar with them.

 

On 4/9/2019 at 2:43 AM, Dinnerb0n3 said:

He says he needs a new GPU because AMD cards are not well suited for machine learning and he has an Rx480.

RX480 can be used for deep learning. There is a framework called ROCm. But it is not the first choice.