Hardware best practices

Talk about Hardware used for Deep Learning
User avatar
bryanlyon
Site Admin
Posts: 601
Joined: Fri Jul 12, 2019 12:49 am
Answers: 40
Location: San Francisco
Has thanked: 3 times
Been thanked: 156 times
Contact:

Re: Hardware best practices

Post by bryanlyon »

I'd switch to the 2060. With the ability to turn on Mixed Precision you'll get huge speed increases and use less ram so you might even get a larger BS overall.


User avatar
stepping-raz0r
Posts: 2
Joined: Fri Dec 11, 2020 3:38 pm
Has thanked: 2 times

Re: Hardware best practices

Post by stepping-raz0r »

Could you give an example of a hardware specs combo in the $2K (+/- $500) price range for someone just starting out?

(Based solely on faceswap considerations...)

Last edited by stepping-raz0r on Fri Dec 11, 2020 4:15 pm, edited 1 time in total.

User avatar
bryanlyon
Site Admin
Posts: 601
Joined: Fri Jul 12, 2019 12:49 am
Answers: 40
Location: San Francisco
Has thanked: 3 times
Been thanked: 156 times
Contact:

Re: Hardware best practices

Post by bryanlyon »

I'm not going to give specific advice since right now supply is difficult and sales/availability/driver support are in constant flux. Instead I'll just say buy the best hardware you can with a particular focus on the GPU. Faceswap can do multiple GPUs even if SLI/NVLink isn't supported. Others have posted requests for info on specific builds and you're likely to get a better response if you request help on a build instead of asking for us to design one for you.


User avatar
abigflea
Posts: 176
Joined: Sat Feb 22, 2020 10:59 pm
Answers: 2
Has thanked: 20 times
Been thanked: 50 times

Re: Hardware best practices

Post by abigflea »

stepping-raz0r wrote: Fri Dec 11, 2020 3:57 pm

Could you give an example...
(Based solely on faceswap considerations...)

Will say Nvidia whatever you do.
30x0 series isn't working quite yet. Nvidia drivers and some other software are not updated to support FS(among other things), but may in a couple months.

:o I dunno what I'm doing :shock:
RTX 3090 : RTX 3070 : RTX 3060 : RTX: 2070 : Ghetto 1060


User avatar
stepping-raz0r
Posts: 2
Joined: Fri Dec 11, 2020 3:38 pm
Has thanked: 2 times

Re: Hardware best practices

Post by stepping-raz0r »

Thanks, I realize there are far too many variables. Sorry, total noob: it's just that I'm looking for something I can unbox and plug in. Would the below nicely get the job done?

AMD Ryzen 7 3800X 3.9GHz 8-Core, B450 MATX, GeForce RTX 2080 Ti 11GB, 16GB DDR4, 960GB SSD


User avatar
bryanlyon
Site Admin
Posts: 601
Joined: Fri Jul 12, 2019 12:49 am
Answers: 40
Location: San Francisco
Has thanked: 3 times
Been thanked: 156 times
Contact:

Re: Hardware best practices

Post by bryanlyon »

That'd be fine, and be quite usable for other uses like gaming. If you can wait for availability (and support) a 3080 might be a good choice instead of the 2080 ti.


User avatar
abigflea
Posts: 176
Joined: Sat Feb 22, 2020 10:59 pm
Answers: 2
Has thanked: 20 times
Been thanked: 50 times

Re: Hardware best practices

Post by abigflea »

JohnnyMacIII wrote: Thu Feb 27, 2020 4:36 am

Thanks for the write up. How much RAM would you recommend? How much of a performance upgrade would it be from a 3700 to a 3900 if running two rtx 2070 supers? Thanks!

John

I think 16Gb would be fine, but I have 32 GB myself.
FaceSwap doesn't require much, but peripheral software like video editors, encoders, and whatnot tend to come into play with doing swaps and they do eat memory.

For your 2x 2070 I would highly suggest that you use Linux for the benefit. Xubuntu, or Ubuntu would be fine.

Additional considerations for your hardware is a good mainboard. Be 100% sure that the 2 slots used for your cards will both run at PCIe 8x for both. If you are really serious, get one that supports x16 for both, but those are costly. Investigate those well.
The GPU do need to speak to each other, and anything less than 8X, you'll lose performance.

My 2x 2070 go about 80% faster than a single.

:o I dunno what I'm doing :shock:
RTX 3090 : RTX 3070 : RTX 3060 : RTX: 2070 : Ghetto 1060


User avatar
RavingEnd
Posts: 2
Joined: Thu Jan 14, 2021 2:08 pm
Has thanked: 1 time

Re: Hardware best practices

Post by RavingEnd »

Hi, I'm not sure if this is the correct thread, but this is hardware related.
I see people in other threads getting 300 EG/s and the like with a 2070 and whatnot.
Using a GTX1060 6GB with Realface seems to net an average of ~35EG/s.
Obviously I'm not expecting anything close to 2070 speeds, but is 35EG/s normal for a 1060?


User avatar
bryanlyon
Site Admin
Posts: 601
Joined: Fri Jul 12, 2019 12:49 am
Answers: 40
Location: San Francisco
Has thanked: 3 times
Been thanked: 156 times
Contact:

Re: Hardware best practices

Post by bryanlyon »

RavingEnd wrote: Thu Jan 14, 2021 2:12 pm

Hi, I'm not sure if this is the correct thread, but this is hardware related.
I see people in other threads getting 300 EG/s and the like with a 2070 and whatnot.
Using a GTX1060 6GB with Realface seems to net an average of 35EG/s.
Obviously I'm not expecting anything close to 2070 speeds, but is 35EG/s normal for a 1060?

Yes, this is normal. It really depends on the model as different models have different amount of computation to be done. Original on a 2070 can't be compared to Realface on a 1060. You're changing too many variables simultaneously.


User avatar
RavingEnd
Posts: 2
Joined: Thu Jan 14, 2021 2:08 pm
Has thanked: 1 time

Re: Hardware best practices

Post by RavingEnd »

bryanlyon wrote: Thu Jan 14, 2021 3:55 pm
RavingEnd wrote: Thu Jan 14, 2021 2:12 pm

Hi, I'm not sure if this is the correct thread, but this is hardware related.
I see people in other threads getting 300 EG/s and the like with a 2070 and whatnot.
Using a GTX1060 6GB with Realface seems to net an average of ~35EG/s.
Obviously I'm not expecting anything close to 2070 speeds, but is 35EG/s normal for a 1060?

Yes, this is normal. It really depends on the model as different models have different amount of computation to be done. Original on a 2070 can't be compared to Realface on a 1060. You're changing too many variables simultaneously.

Many thanks, so it's basically down to playing around with the different models and picking a good balance between speed and desired quality outcome.


User avatar
tokafondo
Posts: 31
Joined: Mon Dec 16, 2019 1:43 pm
Has thanked: 10 times
Been thanked: 4 times

Re: Hardware best practices

Post by tokafondo »

Ok, so RTX Titan, the recommended not to buy card, has 24GB of RAM and 576 tensor cores.

What can be done with it? I mean... Is there actually a benefit of thas monster specs? Can some one train 1024x1024 there with batch sizes of 128? And I'm just putting high numbers there as an exaggeration.

On the practical side, the question would be: If I buy not one but two of these cards, and train 256x256 with batch sizes of 32... Would I get a zillion EG/s and train in 48 hours, what others do train in ten days?

Please help me understand. Thanks.


User avatar
bryanlyon
Site Admin
Posts: 601
Joined: Fri Jul 12, 2019 12:49 am
Answers: 40
Location: San Francisco
Has thanked: 3 times
Been thanked: 156 times
Contact:

Re: Hardware best practices

Post by bryanlyon »

RTX Titan has no real advantages over a 2080 ti which shares a nearly identical chip. It also barely has an advantage against a 2060 ti. Thats why it's not a recommended card.

If you buy 2 of them you'll get training about 2x a 2060 ti at 10x the price. If you really want that much power, just buy 2x 2060 tis which ALSO get nearly 2x a single 2060 ti.


User avatar
cosmico
Posts: 94
Joined: Sat Jan 18, 2020 6:32 pm
Has thanked: 13 times
Been thanked: 28 times

Re: Hardware best practices

Post by cosmico »

abigflea wrote: Fri Dec 11, 2020 10:42 pm

Will say Nvidia whatever you do.
30x0 series isn't working quite yet. Nvidia drivers and some other software are not updated to support FS(among other things), but **may** in a couple months.

Wow, did not know this. My friend was planning on updating his computer and was thinking about getting a rtx 30X0 series. Has there been any updates on if rtx 30X0 series are working yet?


User avatar
torzdf
Posts: 1305
Joined: Fri Jul 12, 2019 12:53 am
Answers: 121
Has thanked: 42 times
Been thanked: 259 times

Re: Hardware best practices

Post by torzdf »

My word is final


Post Reply