NVIDIA 3000 series?

josh

Active Member
Oct 21, 2013
514
136
43
How do the NVIDIA 3000 series perform with ML compared to 2000? Currently using 2070S just purchased early this year. Don't want to upgrade if it's not necessary.
 

dandanio

Active Member
Oct 10, 2017
155
53
28
The question is then: how fast do you need them to perform? Is your current hardware holding you back? Do you feel like 10-20-50% speedup is necessary for you to be more productive? If so, you may need to take a look at 3090 or some such. They are faster, but not really like 50% faster...
 

rafale77

Member
Sep 28, 2020
30
11
8
Don't mean to pick on you but most ML usually doesn't require much GPU. Deep learning does. If you run deep learning frameworks then the question indeed is how much of it do you need? I am also looking at it from a power consumption perspective and though you may not need it to be faster, the 3000 series is definitely a lot more power efficient. So if you are using your GPU a lot today, you may still benefit from some lower power savings. The downside though is that it likely will require some time for the driver and for cuda/cudnn to make full use of the new GPU.
 

josh

Active Member
Oct 21, 2013
514
136
43
Sorry, I totally meant deep learning not ML. I've been experimenting with neural nets but nothing too heavy yet to be frustrated with 2070S. The driver issue does make sense and power savings is definitely a consideration for me.
 

rafale77

Member
Sep 28, 2020
30
11
8
I am on the same boat and am looking to see how the 3070 will perform. My 2070 which I got for a relative bargain is a champ but consumes 60-100W on a constant basis...
 

Dreece

Active Member
Jan 22, 2019
506
154
43
William is going to do the RTX 3090 first. I got confirmation ours is finally getting here next week.
Could you ask William if he can do some kind of performance test with Adobe Premiere comparing say the last gen 2080rtx to the new behemoth 3090?... layer up a load of 4k/8k clips and add some GPU accelerated effects etc and roll out an export on both cards? - real life usage rather than synthetics. Not quite ML/DL, but would be helpful.
 

BackupProphet

Well-Known Member
Jul 2, 2014
824
295
63
Stavanger, Norway
kingmakers.no
I do some deep learning, and I ordered a 3090 RTX just because that my 2080 with 8GB is not enough for some large models which may require 16-20GB of memory. Ironically this is related to NLP stuff, not image classification stuff.

But for the last few years, I've seen a more of a significant performance jump by upgrading libraries. I think Tensorflow 1.2 to 2.3 was like 5-10 times faster training on my CPU. And the 2080 is somewhere between 10-50x faster again
 
  • Like
Reactions: Patrick and T_Minus

Patrick

Administrator
Staff member
Dec 21, 2010
12,111
5,133
113
I do some deep learning, and I ordered a 3090 RTX just because that my 2080 with 8GB is not enough for some large models which may require 16-20GB of memory. Ironically this is related to NLP stuff, not image classification stuff.
I still want to know where to order these!
 

josh

Active Member
Oct 21, 2013
514
136
43
I do some deep learning, and I ordered a 3090 RTX just because that my 2080 with 8GB is not enough for some large models which may require 16-20GB of memory. Ironically this is related to NLP stuff, not image classification stuff.

But for the last few years, I've seen a more of a significant performance jump by upgrading libraries. I think Tensorflow 1.2 to 2.3 was like 5-10 times faster training on my CPU. And the 2080 is somewhere between 10-50x faster again
Have you tried NVLink with 2x 2080?
 

josh

Active Member
Oct 21, 2013
514
136
43
Should I get another 2070S and NVLink with my old card or straight up buy a 2080S? EVGA B-Stock has some pretty amazing prices during the sales.
 

wildpig1234

Well-Known Member
Aug 22, 2016
1,854
284
83
46
another thing is availability. Good luck getting ANY 3000 at retail price. Nvidia and their OEMs sold out their entire stock of 3070, 80, 90 for 2020.
 

josh

Active Member
Oct 21, 2013
514
136
43
another thing is availability. Good luck getting ANY 3000 at retail price. Nvidia and their OEMs sold out their entire stock of 3070, 80, 90 for 2020.
That's not true. I just got a 3070 at msrp via walk in today. Plenty of 3070 stocks and some 3080s reserved for full systems. It's only the online inventories that have been sold out because of the bots.
 
Last edited:

wildpig1234

Well-Known Member
Aug 22, 2016
1,854
284
83
46
That's not true. I just got a 3070 at msrp via walk in today. Plenty of 3070 stocks an some 3080s reserved for full systems. It's only the online inventories that have been sold out because of the bots.
which store did you find it at?
 

wildpig1234

Well-Known Member
Aug 22, 2016
1,854
284
83
46
Does anyone know if my E5-2650V2 and DDR3 will bottleneck the 3070?
depends on what resolution and game. For MSFS 2020 the cpu is the major limitation a lot of time. for other games, if you run at less than 1440p in the 3070 then the cpu could start to be the bottleneck.
 

josh

Active Member
Oct 21, 2013
514
136
43
depends on what resolution and game. For MSFS 2020 the cpu is the major limitation a lot of time. for other games, if you run at less than 1440p in the 3070 then the cpu could start to be the bottleneck.
Not using it for gaming, just deep learning
 

wildpig1234

Well-Known Member
Aug 22, 2016
1,854
284
83
46
The word now is that even though amd navi will be more available, it is also still predicted to be sold out in minutes just like nvidia. I wonder why this is the case with these cards lately. Why the sudden demand/supply imbalance? It can't be all cryptominers or deep learners or is it?

Is the cause being there is more absolute number of people now suddenly buying or is it because they just can't make as many cards nowaday as they used to be?
 

alex_stief

Active Member
May 31, 2016
718
230
43
36
My theory: with RTX 2000 cards, there was no real improvement in performance per dollar. The cards were faster than the 1000 series, but Nvidia also jacked up the prices.
Now with 3000 series being both faster and cheaper than 2000 series, is looks like exceptionally good value. Everyone seems to go along with that perception, when in fact we are just back to the regular generational improvement that would have happened without 2000 series.
Adding to that, many people skipped 2000 series due to the poor value proposition, and are now in "desperate" need of an upgrade to their 1000 series card.
 
Last edited: