RTX3080 - Page 16

James Heinrich · 2020-10-05, 10:42

Quote:

Originally Posted by LaurV

Sometime ago I read on the web about them being 7nm

AMD's RDNA2 / Big Navi / RX6xxx will be on TSMC 7nm.
There's some stories (not sure how true) that NVIDIA was trying to negotiate with TSMC and threatening "if you don't give us a good deal we can always go to Samsung", and TSMC said "ok, bye!" and NVIDIA was stuck with an (arguably worse) fab process. It may (partially) account for the envelope-pushing TDPs of the RTX 3xxx series.

LaurV · 2020-10-05, 11:00

Well, I don't know about those, but if you look to the web "technical" sites, mid-end July, even August, all were talking (speculating??) about 7nm, and only later the "official scores" came out, enough for me to come here and make a fool of myself... haha.

Anyhow, hope for the best, in any case, either more and cheaper 2080s or faster 3080s. Deeper into it, this site shows why one is better than the other (you can click on the cards on the title of the table to reverse the direction of the comparison). Lower memory bandwidth, as well as speed, is for sure some handicap for FFT. They may be, however, much faster for TF, as the core clock is much faster and there are more cores. If (and I repeat, IF) I should be able to run my 2080Ti's to 320W without smoking them hard, I don't see much difference either, they would come out even. Deeper and deeper I go into it, I believe more and more that this is (again! from NV) just an intermediary step. We may witness some bigger silicon coming out soon. But OTOH, this may be only me being paranoid.

M344587487 · 2020-10-06, 19:30

Phoronix compute benchmarks: https://www.phoronix.com/scan.php?pa...-compute&num=1

LaurV · 2020-10-07, 04:19

Very comprehensive, thank you for sharing.

Page 8 is really interesting for our common interest and also page 3 in the middle, where a DPFP (FP64) graph is shown. You can witness a 29% higher FFT performance of 3080, but let's not forget that the 3080 takes exactly 29% more power! So, confirming my assumptions, if I could run my 2080Ti's at 320W instead of 250W, without smoking them.... (actually, when FP32 is used, I can get close to 280W for the two mix-cooled cards I have).

Talking about FP32, I have to confess that Page 2 put me off a bit when I saw such a low performance of 2080Ti and Titans at "fold-it". I use to "fold" from time to time and my cards' performances are way better. I mean WAY better (and I can run them close to 280 watts, as said). I almost jumped to the ceiling "what a bullshit?", then I read the "per dollar" part of the title of the graphic

They consider the Ti at $1200 when you can actually buy it way below $800 (but say 800 for easy calculus). Therefore all numbers for Ti there have to be multiplied by 3/2, while, by the same logic, all numbers for 3080 have to be divided by some "over-unity" constant (who the hack sells 3080 for $700 now?? And you - generic you - will buy it much higher soon!)

jvang · 2020-10-07, 17:48

Quote:

Originally Posted by M344587487

Phoronix compute benchmarks: https://www.phoronix.com/scan.php?pa...-compute&num=1

lots of these compute benchmark roundups with the 3080/3090 leave out the 1080ti for some reason...

Quote:

Originally Posted by LaurV

You can witness a 29% higher FFT performance of 3080, but let's not forget that the 3080 takes exactly 29% more power! So, confirming my assumptions, if I could run my 2080Ti's at 320W instead of 250W, without smoking them.... (actually, when FP32 is used, I can get close to 280W for the two mix-cooled cards I have).

They consider the Ti at $1200 when you can actually buy it way below $800 (but say 800 for easy calculus). Therefore all numbers for Ti there have to be multiplied by 3/2, while, by the same logic, all numbers for 3080 have to be divided by some "over-unity" constant (who the hack sells 3080 for $700 now?? And you - generic you - will buy it much higher soon!)

i've been thinking similarly, ampere seems to be the same performance per watt as turing in a lot of applications which is a bit strange to me. if turing cards hadnt been so overpriced and provided such small gains over pascal (barring tensor core/rtx core stuff), ampere wouldnt look so spectacular by comparison

firejuggler · 2020-10-07, 19:04

semi related https://www.techradar.com/news/nvidi...smoothly-at-4k

pinhodecarlos · 2020-10-22, 11:43

Stay tuned for SRBase fórmula BOINC sprint which will bring more GPU power from 23rd to 26th. This might mean this nvidia cards series appearing on the list but most importante more TF work done. I can see more than 400k wus on the queue, big dump is expected.

NoahGreat · 2020-11-06, 13:42

Quote:

Originally Posted by M344587487

Phoronix compute benchmarks: https://www.phoronix.com/scan.php?page=article worktime &item=nvidia-rtx3080-compute&num=1

Hope to get one of these QT`ies in December...

M344587487 · 2020-11-06, 18:10

Quote:

Originally Posted by NoahGreat

Hope to get one of these QT`ies in December...

If you're waiting that long then the upcoming AMD 6000 series might change your plans. The 6800 and 6800XT are released in 12 days and look to be competitive at least in gaming, how that shakes out for us remains to be seen: https://mersenneforum.org/showthread.php?t=26134

2020-10-05, 11:00	#167
LaurV Romulan Interpreter "name field" Jun 2011 Thailand 3·23·149 Posts	Well, I don't know about those, but if you look to the web "technical" sites, mid-end July, even August, all were talking (speculating??) about 7nm, and only later the "official scores" came out, enough for me to come here and make a fool of myself... haha. Anyhow, hope for the best, in any case, either more and cheaper 2080s or faster 3080s. Deeper into it, this site shows why one is better than the other (you can click on the cards on the title of the table to reverse the direction of the comparison). Lower memory bandwidth, as well as speed, is for sure some handicap for FFT. They may be, however, much faster for TF, as the core clock is much faster and there are more cores. If (and I repeat, IF) I should be able to run my 2080Ti's to 320W without smoking them hard, I don't see much difference either, they would come out even. Deeper and deeper I go into it, I believe more and more that this is (again! from NV) just an intermediary step. We may witness some bigger silicon coming out soon. But OTOH, this may be only me being paranoid. Last fiddled with by LaurV on 2020-10-05 at 11:00

2020-10-07, 04:19	#169
LaurV Romulan Interpreter "name field" Jun 2011 Thailand 3·23·149 Posts	Very comprehensive, thank you for sharing. Page 8 is really interesting for our common interest and also page 3 in the middle, where a DPFP (FP64) graph is shown. You can witness a 29% higher FFT performance of 3080, but let's not forget that the 3080 takes exactly 29% more power! So, confirming my assumptions, if I could run my 2080Ti's at 320W instead of 250W, without smoking them.... (actually, when FP32 is used, I can get close to 280W for the two mix-cooled cards I have). Talking about FP32, I have to confess that Page 2 put me off a bit when I saw such a low performance of 2080Ti and Titans at "fold-it". I use to "fold" from time to time and my cards' performances are way better. I mean WAY better (and I can run them close to 280 watts, as said). I almost jumped to the ceiling "what a bullshit?", then I read the "per dollar" part of the title of the graphic They consider the Ti at $1200 when you can actually buy it way below $800 (but say 800 for easy calculus). Therefore all numbers for Ti there have to be multiplied by 3/2, while, by the same logic, all numbers for 3080 have to be divided by some "over-unity" constant (who the hack sells 3080 for $700 now?? And you - generic you - will buy it much higher soon!) Last fiddled with by LaurV on 2020-10-07 at 06:40

2020-10-06, 19:30	#168
M344587487 "Composite as Heck" Oct 2017 3·311 Posts	Phoronix compute benchmarks: https://www.phoronix.com/scan.php?pa...-compute&num=1

2020-10-07, 19:04	#171
firejuggler "Vincent" Apr 2010 Over the rainbow 101101000100₂ Posts	semi related https://www.techradar.com/news/nvidi...smoothly-at-4k

2020-10-22, 11:43	#172
pinhodecarlos "Carlos Pinho" Oct 2011 Milton Keynes, UK 7·733 Posts	Stay tuned for SRBase fórmula BOINC sprint which will bring more GPU power from 23rd to 26th. This might mean this nvidia cards series appearing on the list but most importante more TF work done. I can see more than 400k wus on the queue, big dump is expected.