GPU for machine learning, physics simulations and gaming

[alkakxmmiq]

Hi!

 

I am thinking about buying a new PC but can't decide which GPU is the best bang for buck for me. I dream of a Titan V which is hands down the best for my purposes (tensor cores + 1/2 fp64 FLOPS compared to fp32 FLOPS is a killer feature for computer modelling in physics since normally nVidia cards have 1/32; sadly they disabled ECC) but it is probably unwise to spend that much money since I'm just starting.

 

So I'm going mainly for bang for buck.

 

The main contestants are RTX 2070 or second hand either 1080ti, Titan Z, titan black or perhaps some others.

 

Titan Z has great fp64 performance, not bad fp32 performance but has a really high TDP and an older architecture and can be purchased for the lower end of the price of the second hand 1080ti.

 

Titan Black is really cheap and has pretty good fp64 performance, otherwise it's quite old.

 

1080ti is all round great (memory, bandwidth, fp32 performance) but does not have Tensor cores (does anyone have experience with how tensor cores boost performance in ML compared to normal parallelization?), not that much performance on fp64 and has higher TDP than 2070.

 

rtx 2070 does not have the most memory, memory bandwidth or fp32 performance, but all of it is decent. fp64 performance is not that great but it has Tensor cores (again, does anyone have experience with them in ML applications?), great fp16 performance (idk how the older ones fare on fp16), DLSS and HW ray tracing for games or graphics, new architecture and good performance/TDP ratio.

 

other considerations:

I've looked at eBay and they have used tesla k80s for about 650 USD. Is that possible? If seems like a great deal since k80 has ECC, great fp64 performance and lots of memory. Sadly no fp16 performance boost like on Volta and no tensor cores but the ECC is still really valuable for simulations (not that much for ML afaik), also I would have to use something else for gaming and other kinds of work since it does not have a display output (is it possible to wing it somehow?).

 

tl;dr: Do you have experience with workloads mentioned in the title? Which GPUs would you recommend for them taking into account bang for buck performance.

[XR6]

I'd go for the 1080 Ti.

[Fasauceome]

1 minute ago, 1kv said:

I'd go for the 1080 Ti.

There are tensor cores on the 2070, I believe that gives it a serious edge

[Radium_Angel]

6 minutes ago, alkakxmmiq said:

experience

I do machine learning, specifically RNN, Recurrent Neural Networks. I've worked on both AMD and nVidia GFX cards.

Skip AMD, the software isn't mature on RNN...yet.

I use a GTX 660 and the iterations for char-rnn and word-rnn run sub second (I want to say about 1/3 of a second? I don't recall offhand right now) under Ubuntu 14.04LTS.

 

GTX660s are dirt cheap, and a good way to start if you are unsure if you want to burn more money immediately, or ease into ML.

Just an option...

[alkakxmmiq]

1 minute ago, 1kv said:

I'd go for the 1080 Ti.

I hope there will be plenty available now after Christmas So I'm keeping my eye on them. It's great that top of the line 1080ti is not that more expensive than a new 2070. The real question is: Can I live without raytracing? 

[XR6]

1 minute ago, alkakxmmiq said:

I hope there will be plenty available now after Christmas So I'm keeping my eye on them. It's great that top of the line 1080ti is not that more expensive than a new 2070. The real question is: Can I live without raytracing? 

Raytracing isn't needed for gaming but as @fasauceome said the tensor cores on the RTX 2070 will probably give it an edge so I would go for the 2070 instead.

 

 

[alkakxmmiq]

Just now, Radium_Angel said:

I do machine learning, specifically RNN, Recurrent Neural Networks. I've worked on both AMD and nVidia GFX cards.

Skip AMD, the software isn't mature on RNN...yet.

I use a GTX 660 and the iterations for char-rnn and word-rnn run sub second (I want to say about 1/3 of a second? I don't recall offhand right now) under Ubuntu 14.04LTS.

 

GTX660s are dirt cheap, and a good way to start if you are unsure if you want to burn more money immediately, or ease into ML.

Just an option...

Thanks for the experience! I am thinking more about the Titan Blacks in this category since while searching local second hand deals I have deals which are about 40 % more expensive than 660s but it has incomparable performance on the double precision calculations which is crucial for finite element method calculations, n-body simulations and such.

 

On ML I think one would rather have half-precision with greater speeds, that's why I'm asking for someone's experience on the RTX cards.

 

Anyway thanks for the input, it's great to have some real-world data so I can get a feel for how long it will run these tasks.

[fabafaba]

28 minutes ago, alkakxmmiq said:

I hope there will be plenty available now after Christmas So I'm keeping my eye on them. It's great that top of the line 1080ti is not that more expensive than a new 2070. The real question is: Can I live without raytracing? 

The 1080ti is perfectly fine for Tensorflow, so go get one!  If you're curious about the performance compared to other GPUs, you can read Don from Puget Systems article on that: https://www.pugetsystems.com/labs/hpc/RTX-2080Ti-with-NVLINK---TensorFlow-Performance-Includes-Comparison-with-GTX-1080Ti-RTX-2070-2080-2080Ti-and-Titan-V-1267/

[alkakxmmiq]

21 minutes ago, fabafaba said:

The 1080ti is perfectly fine for Tensorflow, so go get one!  If you're curious about the performance compared to other GPUs, you can read Don from Puget Systems article on that: https://www.pugetsystems.com/labs/hpc/RTX-2080Ti-with-NVLINK---TensorFlow-Performance-Includes-Comparison-with-GTX-1080Ti-RTX-2070-2080-2080Ti-and-Titan-V-1267/

But it has only 1/32 double precision performance compared to its single precision (32bit float) one. The Titans and such have modes where you can get 1/3 (or 1/2 in the case of Titan V) double precision performance compared to the regular precision. I'm on odds with the 1080ti for that reason

 

There are so many variables, there is no clear winner and I'm a really indecisive person

[alkakxmmiq]

Just now, alkakxmmiq said:

But it has only 1/32 double precision performance compared to its single precision (32bit float) one. The Titans and such have modes where you can get 1/3 (or 1/2 in the case of Titan V) double precision performance compared to the regular precision. I'm on odds with the 1080ti for that reason

 

Basically there are so many variables, no clear winner and I'm a really indecisive person