• Guest, The rules for the P & N subforum have been updated to prohibit "ad hominem" or personal attacks against other posters. See the full details in the post "Politics and News Rules & Guidelines."

Question The Ampere architecture is designed more towards compute task vs sheer gaming?

traderjay

Senior member
Sep 24, 2015
206
150
116
Forgive me if this has been asked before but I read somewhere that the Ampere architecture is designed more towards compute task vs sheer gaming compared to turing? This has something to do with how the FP32 is executed in the new architecture? Can someone please point me towards some sources for reading/info please? My RTX 3090 is primarily used for work (video editing) and some gaming (MSFS 2020 and Arma 3) so it will be interesting to see how future non-gaming application can take advantage of this architecture.
 

Markfw

CPU Moderator, VC&G Moderator, Elite Member
Super Moderator
May 16, 2002
20,810
8,970
136
Forgive me if this has been asked before but I read somewhere that the Ampere architecture is designed more towards compute task vs sheer gaming compared to turing? This has something to do with how the FP32 is executed in the new architecture? Can someone please point me towards some sources for reading/info please? My RTX 3090 is primarily used for work (video editing) and some gaming (MSFS 2020 and Arma 3) so it will be interesting to see how future non-gaming application can take advantage of this architecture.
Install and run folding@home. You have to get a passkey to get full credit, then do 10 units. I would like to know the PPD,. Right now my 2080TI's are doing 5 million ppd. If the 3090 will do 10-15 million ppd, then I would get one. Make sure you use team 198 though.
 

Guru

Senior member
May 5, 2017
809
312
106
Anandtech used to do a great analysis of architecture when they actually did reviews, but nowadays there is barely any actual competent journalists who know what they are talking about, its all pundits and enthusiasts, but not actual hardware engineers or phd's. Trash articles, rumors, exuberant claims bring in more viewers and thus more money, so who would hire an actual tech guru and pay them top money for savvy articles when barely half their audience reads it, so they hire literally anybody to just recycle the latest news of the day and post rumors!
 
  • Like
Reactions: xpea

Kenmitch

Diamond Member
Oct 10, 1999
8,152
1,592
126
Install and run folding@home. You have to get a passkey to get full credit, then do 10 units. I would like to know the PPD,. Right now my 2080TI's are doing 5 million ppd. If the 3090 will do 10-15 million ppd, then I would get one. Make sure you use team 198 though.
You can see some results in the following links. I'm not into folding@home, but it looks like it wouldn't be worth the investment. Probably 2 3080's or possibly RDNA2 would be better? Of course it's probably best to just ride out the Ti's in the end.


https://www.reddit.com/r/Folding/comments/imfuxd
 

Markfw

CPU Moderator, VC&G Moderator, Elite Member
Super Moderator
May 16, 2002
20,810
8,970
136
You can see some results in the following links. I'm not into folding@home, but it looks like it wouldn't be worth the investment. Probably 2 3080's or possibly RDNA2 would be better? Of course it's probably best to just ride out the Ti's in the end.


https://www.reddit.com/r/Folding/comments/imfuxd
I don't believe that. Right now I have 2 2080TI's that do their work units in 40 seconds and are running at 5 million ppd. The average there was 5 million at 69 seconds on a 3080 !! The 2080TI has 4352, the 3080 has 8704 and the 3090 has 10496 ! The cuda cores make a big difference in F@H PPD, so right now I would trust @traderjay 's numbers if he did this.

The 3080 should be about 10 million ppd with bonus, and the 3090 should be close to 15 million.

It is possible that since they may not have beta enabled, its used the old methodology, but I was only getting 2.5 to 3.6 million ppd then, so it could be right.
 

jimhsu

Senior member
Mar 22, 2009
705
0
76
The Ampere series is quite impressive for DL performance (of course, AMD lacking a coherent strategy in this space, and ROCm not "quite" working on even Navi just yet, etc etc). Performance scaleups more than potentially 2x 2080 ti makes the 3080 possibly the best buy in the space, even compared to multiple used cards, and the 3090 not a bad buy either (highest performance/dollar leading 24GB card in the graph below). Personally eyeing the 3090 purely for transformer NLP model training and image GANs, which take 24GB, 48GB, basically any amount of memory you can throw at it. More memory > larger batch sizes > better convergence. The gaming part is simply not interesting in comparison, and Navi probably worth waiting for that.

1601759030476.png

 

Qwertilot

Golden Member
Nov 28, 2013
1,500
179
106
Forgive me if this has been asked before but I read somewhere that the Ampere architecture is designed more towards compute task vs sheer gaming compared to turing? This has something to do with how the FP32 is executed in the new architecture? Can someone please point me towards some sources for reading/info please?
As above, unsure of good sources. The base architecture itself is, of course, aiming to balance both gaming and compute tasks. That's always been true I guess, but explicitly so ever since they put tensor cores into a lot of the chips back in the last generation & started getting such big sales for deep learning etc.

It seems to do that basic goal rather well. Then its a question of what they do with it.

A100 is obviously a compute only part.

All the other parts are going to be a mixture. Its seems fair to say that A102 was designed with a non trivial focus on compute uses, with gaming as a major byproduct. A few hints there but basically if it had been designed for gaming first it'd have been a bit smaller, likely less power draw.

The huge power draw (no laptops!), there's apparently a little bit more 'raw' FP32 than it can use in games, the huge amount of vram on the 3090 etc.

The 3070 down is seemingly going into laptops as a mobile part so it might well be balanced as a much purer gaming chip. I presume the less powerful chips are just less interesting for workstations anyway.
 

CP5670

Diamond Member
Jun 24, 2004
4,485
58
91
The 3090 is a great deep learning card because of its 24GB memory, but it's worth comparing it to the perf/dollar on a cloud service. This has increasingly become the standard way to train neural networks these days instead of buying local machines with video cards for doing it.
 

jimhsu

Senior member
Mar 22, 2009
705
0
76
The 3090 is a great deep learning card because of its 24GB memory, but it's worth comparing it to the perf/dollar on a cloud service. This has increasingly become the standard way to train neural networks these days instead of buying local machines with video cards for doing it.
Good point. For production workloads, the CUDA license prohibits most companies from using consumer grade hardware anyways. But for research/hobbyist use, datacenter/cloud doesn’t make a lot of sense if your utilization is above 10% or so (based on current AWS prices, with 1yr amortization, I don’t recall the specific details)
 

ASK THE COMMUNITY