GPU Benchmarks

[houding]

In the GPU "benchmarks", the 580 is rated higher than most other GPUs (680, 670, 770, 760 - I looked at these 5).

But looking at other benchmark sites, the 580 is rated the lowest.

Is there a difference between gaming performance and "sieving" performance?

[kracker]

Yes. Kepler performs lower in compute than Fermi.

[tului]

Something about crippled double precision floats to differentiate if I'm not mistaken. It's why I only use AMD(or at least did for Bitcoins)

[kracker]

Originally Posted by tului

Something about crippled double precision floats to differentiate if I'm not mistaken. It's why I only use AMD(or at least did for Bitcoins)

That's for LL tests. TF is almost all if not all single precision.

[TheJudger]

TF is all about integer throughput (at least for mfakt[co]).

[TheMawn]

I think there's something not quite right between mfaktc and the Kepler architecture. I don't have the knowledge or skills to back this claim up or do anything about it, but I just can't believe that a GTX 560 Ti is better at this than a GTX 780.


Anyway, as it stands, yes, the GTX 580 is one of the strongest cards for compute. There are some economical purchases to be made. I look these up on eBay once in a while just for fun. Among the blithering idiots who still think their heavily used three-year-old cards are worth 90% of their initial retail value are some more sensible people starting bids at $100-$150 for a GTX 580. With some patience, a person could populate a few extra PCI-E slots with GTX 580's for $150 apiece.

On the other hand, we're not the only people who know that the GTX 580 was a god of computing, so there are other people out there looking (which is why a used GTX 580 isn't straight-up free) and a lot of the cards are themselves retired from heavy compute environments. Some of the cards look like they're in terrible shape. I believe LaurV is one individual who has had some success with purchasing used high-performance GPU's but at least on eBay Canada, it seems there are more people trying to squeeze money out of their junk than anything else.

I didn't see it today but a few weeks ago, someone was trying to sell their broken GPU. "One spot is burnt and a couple of capacitors exploded but I know some people are into repairs and stuff so I'm putting it up for them if they want to try to fix it."

$100 price tag.

[LaurV]

There is nothing wrong with the software, (some of) the 7xx series are indeed worse than (some of) their 5xx corespondents, and one of the reasons, beside the different architecture, and talking only about integer calculus (like in the case of 780 vs 580) is the power limitation for keplers, they were designed to be more "efficient", as opposed to fermis, which were designed to be powerful. Think about the fact that 780 is power-limited, it can not go higher than its TDP, or about 80% of it, in fact. That is why I am always talking in my former posts about "unlocked Titans", those which you can set the SP or DP mode, and set the amount of power the card burns. Numeric example: a "normal" Titan gets about 400 GHzD/D, and a 580 gets 430 GHzD/D. Unlock the titan and you get 500 GHzD/D.

[fivemack]

Originally Posted by TheMawn

With some patience, a person could populate a few extra PCI-E slots with GTX 580's for $150 apiece.

I've done this (not as many extra PCI-E slots as some, because the GTX580 cards I found turned out to be about 2.5 slots wide); it is worth pointing out that the cards use really significant electricity - 200W per card, which for me amounts to more than the cost of the card every six months, and makes them less power-efficient than CPUs for doing ECM.

(512 stage-1 ECM curves in 20,000 seconds @ 200W = 4MJ on the GPU; 512 stage-1 ECM curves in 4,000 seconds @ 550W = 2.2MJ on 48-core Opteron machine)

I use the GPUs anyway, because the Opteron machine cost $7,000 and I'm not going to buy another one any time soon ...