RTX3080 - Page 9

kruoli · 2020-09-18, 11:13

They have had severe problems:
http://videocardz.com/newz/nvidia-is...y-out-of-stock

pinhodecarlos · 2020-09-18, 11:13

Following scan.co.uk I can see people complaining about extra charges or even due to lag of website charges up to 3x...

M344587487 · 2020-09-18, 11:34

It's the now standard launch farce. Could be partly down to being a paper launch but the numbers aren't known, it just seems a bigger farce this time as it's the most anticipated card of the year. There has been an amusing push back against scalpers this time, namely people bidding single card auctions up to 5 figures then ghosting the seller.

storm5510 · 2020-09-18, 16:32

Quote:

Originally Posted by jwnutter

So, will the current slate of GPU tools (mfaktc, gpuOwl, CUDALucas, etc.) be ready/capable of utilizing the NVIDIA Ampere series

What is the significance of the Ampere, a.k.a. Amp, series?

pinhodecarlos · 2020-09-18, 16:41

Some benchmarks here from PrimeGrid:

https://www.primegrid.com/forum_thre...ap=true#143387

Viliam Furik · 2020-09-18, 17:08

Quote:

Originally Posted by pinhodecarlos

Some benchmarks here from PrimeGrid:

https://www.primegrid.com/forum_thre...ap=true#143387

Aaaaand, that gives an estimate of about 37 hours on 108M exponent. Compared to my Radeon VIIs 27 hours (1000 mV, -5% power, 1600 MHz core, 1025 MHz mem - Afterburner settings) it's about 72 % of its power. Not bad. My RTX 2080 Ti (90% power, rest is stock) gives about 85 hours (2400 us/it on 93M). 3080 is a bit more than 2 times as powerful in Mersenne testing.

All my calculations are about correct, assuming the numbers on the PrimeGrid forum were test results, not factoring results.

If somebody wants to double-check my numbers, here are used values:

R7 -> 108M - 0,92 ms/iter
2080Ti -> 93M - 2,4 ms/iter
3080 was 15,6 hours on 21449434-digit number, so it's equivalent to 71,2M Mersenne exponent.

henryzz · 2020-09-18, 18:17

Quote:

Originally Posted by M344587487

It's the now-standard launch farce. Could be partly down to being a paper launch but the numbers aren't known, it just seems a bigger farce this time as it's the most anticipated card of the year. There has been an amusing push back against scalpers this time, namely people bidding single card auctions up to 5 figures then ghosting the seller.

From what I have heard not a paper launch. People are buying more currently due to being home because of COVID. Production of previous-gen stopped before launch and I believe stocks were low of them so people waited for the new-gen. It has added up to stocks running out really quick. Initial stocks were at a normal level for a launch in an unusual situation.

moebius · 2020-09-18, 20:28

Quote:

Originally Posted by henryzz

From what I have heard not a paper launch. People are buying more currently due to being home because of COVID.

People are more buying because of the conspiracy of the MOD:{expletive deleted}.... .
I'm sorry but this my oppinion.

MOD: We don't know what you're trying to say, and the language you chose didn't help.

tServo · 2020-09-21, 13:49

WRT the Ampere architecture, most of it seems directed towards the gamers and AI researchers. Still, I found 3 things in the latest release of the v11.0 CUDA toolkit that look VERY interesting:

(1) Asynchronous copy between global and shared memory. Optimizing memory operations is EVERYTHING in GPU coding and memory copying has been a part of almost every well written program. This has the potential to really pay off big time since shared memory access is SO much faster than global.

(2) L2 cache management instructions. Again, paying careful attention to memory access patterns can be very rewarding.

(3) I'm not so sure this will pay off but the warp matrix operations now support FP64. This needs to be approached cautiously since these Tensor core operations are meant for AI stuff which don't have the same accuracy requirements wrt rounding that we do.

Since these features are in the toolkit and not necessarily the Ampere architecture itself, that means they can be retro-fitted to older video boards, perhaps even going back to Pascal?

tServo · 2020-09-21, 23:36

Correction on my prior post.
I think Ampere is required for these nifty new features.

kriesel · 2020-09-21, 23:52

https://www.ebay.com/itm/ZOTAC-GAMIN...d/203110063212

2020-09-21, 13:49	#97
tServo "Marv" May 2009 near the Tannhäuser Gate 2²×3×67 Posts	WRT the Ampere architecture, most of it seems directed towards the gamers and AI researchers. Still, I found 3 things in the latest release of the v11.0 CUDA toolkit that look VERY interesting: (1) Asynchronous copy between global and shared memory. Optimizing memory operations is EVERYTHING in GPU coding and memory copying has been a part of almost every well written program. This has the potential to really pay off big time since shared memory access is SO much faster than global. (2) L2 cache management instructions. Again, paying careful attention to memory access patterns can be very rewarding. (3) I'm not so sure this will pay off but the warp matrix operations now support FP64. This needs to be approached cautiously since these Tensor core operations are meant for AI stuff which don't have the same accuracy requirements wrt rounding that we do. Since these features are in the toolkit and not necessarily the Ampere architecture itself, that means they can be retro-fitted to older video boards, perhaps even going back to Pascal? Last fiddled with by tServo on 2020-09-21 at 13:50

2020-09-21, 23:52	#99
kriesel "TF79LL86GIMPS96gpu17" Mar 2017 US midwest 2⁴×461 Posts	A real deal on an RTX3080. Oh, wait. https://www.ebay.com/itm/ZOTAC-GAMIN...d/203110063212

2020-09-18, 11:13	#89
kruoli "Oliver" Sep 2017 Porta Westfalica, DE 2×11×61 Posts	They have had severe problems: http://videocardz.com/newz/nvidia-is...y-out-of-stock

2020-09-18, 11:13	#90
pinhodecarlos "Carlos Pinho" Oct 2011 Milton Keynes, UK 7·733 Posts	Following scan.co.uk I can see people complaining about extra charges or even due to lag of website charges up to 3x...

2020-09-18, 11:34	#91
M344587487 "Composite as Heck" Oct 2017 3×311 Posts	It's the now standard launch farce. Could be partly down to being a paper launch but the numbers aren't known, it just seems a bigger farce this time as it's the most anticipated card of the year. There has been an amusing push back against scalpers this time, namely people bidding single card auctions up to 5 figures then ghosting the seller.

2020-09-18, 16:41	#93
pinhodecarlos "Carlos Pinho" Oct 2011 Milton Keynes, UK 7×733 Posts	Some benchmarks here from PrimeGrid: https://www.primegrid.com/forum_thre...ap=true#143387

2020-09-21, 23:36	#98
tServo "Marv" May 2009 near the Tannhäuser Gate 2²×3×67 Posts	Correction on my prior post. I think Ampere is required for these nifty new features.