Is duel gpu worth it for tensorflow deep learning?

[bigdog1565]

Hi,

 

I was planning to use two gtx 1660 gpus for deep learning with tensorflow and i just wanted to know is it worth it or should i get a single better gpu?

[Grabhanem]

2 minutes ago, bigdog1565 said:

Hi,

 

I was planning to use two gtx 1660 gpus for deep learning with tensorflow and i just wanted to know is it worth it or should i get a single better gpu?

I'd look at this article:

https://www.pugetsystems.com/labs/hpc/TensorFlow-Performance-with-1-4-GPUs----RTX-Titan-2080Ti-2080-2070-GTX-1660Ti-1070-1080Ti-and-Titan-V-1386/

 

It seems to scale pretty well-- two 1660Tis are close to a 2080, so I wouldn't be surprised if two 1660s is around a 2070 or 2070Super.

 

Another option would be the 2060KO-- it uses a cut-down 2080 die, so it tends to do very well in compute.

[bigdog1565]

ok thanks for the advice

[Mateyyy]

If I had to guess, a single RTX 2060 would smoke the two GTX 1660s simply because it's got actual dedicated hardware for deep learning on it, in the form of Tensor cores.

[PCPartPickerPro9171787]

It depends if tensorflow was designed for dual GPU. I believe it was far-cry 3 that was specifically  designed for dual GPUs, so you had the potential to 100%+ frames rate. Most games, however, are not designed like that, so you get much less than 100 %+. 

[Levent]

4 minutes ago, PCPartPickerPro9171787 said:

It depends if tensorflow was designed for dual GPU. I believe it was far-cry 3 that was specifically  designed for dual GPUs, so you had the potential to 100%+ frames rate. Most games, however, are not designed like that, so you get much less than 100 %+. 

Gonna need source on that (FC3 being designed for dual GPUs part) That era of games were particularly good on SLI. Nvidia made great efforts to make sure because they were still betting big on SLI.

OP: Nowadays I would always get the best single GPU card first then add another if your workload needs it. Single card is always less hassle than dual cards.

Also it's dual not duel.

[PCPartPickerPro9171787]

2 minutes ago, Levent said:

Gonna need source on that (FC3 being designed for dual GPUs part) That era of games were particularly good on SLI. Nvidia made great efforts to make sure because they were still betting big on SLI.

OP: Nowadays I would always get the best single GPU card first then add another if your workload needs it. Single card is always less hassle than dual cards.

Also it's dual not duel.

I might have the game wrong. Point being that it's up to the program designer to design the program in such a way to take advantage of duel GPU s similar to how it is up to the a program designer to design a program to use multiple cores on the CPU.    

[igormp]

22 hours ago, bigdog1565 said:

Hi,

 

I was planning to use two gtx 1660 gpus for deep learning with tensorflow and i just wanted to know is it worth it or should i get a single better gpu?

Do you already have a 1660? If so, then go for it. You actually don't need to use the same card, you can try to get the best one you can afford ATM and mix it with your current 1660. If not, then try to go for a RTX card due to the tensor cores, preferably a 2060 Super/2070 due do the 8gb of vram.

 

22 hours ago, PCPartPickerPro9171787 said:

It depends if tensorflow was designed for dual GPU. I believe it was far-cry 3 that was specifically  designed for dual GPUs, so you had the potential to 100%+ frames rate. Most games, however, are not designed like that, so you get much less than 100 %+. 

 

22 hours ago, Levent said:

Gonna need source on that (FC3 being designed for dual GPUs part) That era of games were particularly good on SLI. Nvidia made great efforts to make sure because they were still betting big on SLI.

OP: Nowadays I would always get the best single GPU card first then add another if your workload needs it. Single card is always less hassle than dual cards.

Also it's dual not duel.

This is not a gaming load, it scales pretty much linearly with more GPUs, even different ones. And such cases 2x GPUs is better than one since you get double the performance and vram. Check out the link posted by @Grabhanem for more conclusive results.

[Levent]

5 minutes ago, igormp said:

 

This is not a gaming load, it scales pretty much linearly with more GPUs, even different ones. And such cases 2x GPUs is better than one since you get double the performance and vram. Check out the link posted by @Grabhanem for more conclusive results.

I am aware of what OP is asking. OP hasn't mentioned that they already owned a 1660. If I knew that I would have suggested something different. 

[igormp]

43 minutes ago, Levent said:

I am aware of what OP is asking. OP hasn't mentioned that they already owned a 1660. If I knew that I would have suggested something different. 

My bad then, I just found it weird that you guys were talking about games and SLI in a thread about tensorflow. I also don't know if OP already has a 1660 or not, that's why I suggested 2 scenarios.