Join PrimeGrid
Returning Participants
Community
Leader Boards
Results
Other
drummers-lowrise
|
1)
Message boards :
General discussion :
BOINC Pentathlon and Manual Credit
(Message 117556)
Posted 1858 days ago by  John P. Myers
Simply put, in the Pentathlon, bunkering is in no way cheating. This is why a project is announced 5 days in advance. Bunkering does not break any rules. It is allowed, encouraged and expected.
This is not true for every competition, but for the Pentathlon, it is.
|
2)
Message boards :
Number crunching :
Kepler rumors
(Message 52059)
Posted 4082 days ago by John P. Myers
NeoMetal* wrote:
The Kepler DP floating point architecture is much different than Fermi's. The GK104 (the 680) is not crippled in the same way as Fermi. Kepler is DESIGNED with a limited number of FP64 CUDA cores.
Yes that is true, but crippled is crippled. The method doesn't matter much :)
And yes, since the GTX680 was originally supposed to have been a 670 Ti, there is a chance that "big kepler" won't be reduced to 1/24th, but either 1/16th or 1/8th instead.
If had to place a bet, i'd go with 1/16th. The reason is because the 560 Ti was reduced to 1/12th and the 680 (really a 670 Ti) was reduced by 1/24th, showing they decided to double the amount of crippling. If that trend continues, then "big kepler" would be 1/16th instead of 1/8th.
Also, i'm not sure how old that article is, but "big kepler" should be GK110, not GK100.
2 years ago when they began reducing the DP speed on the 400 series, we believed it was done in the GPU's BIOS so we tried to hack it in order to free up the processing power. However that wasn't the case. There's nothing that can be done. And now that the 680 is a physical crippling, it just drives that point home.
And don't bother crossing your fingers thinking Nvidia will allow full DP performance on any future GPU. They will never risk losing sales of their Tesla line. Ever. And to be honest i find it quite sad that people are suckered into paying $2500+ for a Tesla, when if they'd just use OpenCL in the first place, they could use that $2500 on AMD GPUs and get over twice the computational DP performance.
|
3)
Message boards :
Number crunching :
Kepler rumors
(Message 52032)
Posted 4082 days ago by John P. Myers
[quote]And the bad news continues:
http://vr-zone.com/articles/nvidia-kepler-gpus--back-to-games-away-from-compute-/15332.html
If the 'Big chip' has a lot more DP cores built in, that then would make it (possibly a GTX 685?) a lot more respectable performance overall, even with OpenCL, but it still won't be as fast as AMD/ATI. Of coarse hand tuned CUDA apps will still be around, but there will be less of them since an OpenCL app is much easier to to build, as some here already know.
The "Big Chip" will have SP and DP performance 50% higher than the 680, assuming the same clock speed. DP will still be crippled at 1/24th so it's still basically worthless except for sieving.
Just so we understand exactly what we're missing out on here, the GTX 680 would have a DP speed of 1545 GFLOPS if uncrippled, which is about 40% faster than an HD 6990. The 580 would have 790.55 GFLOPS DP instead of the 197.64 we're "allowed" to use.
|
4)
Message boards :
Number crunching :
Kepler rumors
(Message 51937)
Posted 4085 days ago by John P. Myers
If that is true, then how come all the cards here are NVidia?
Because i'm only talking about double precision. Single precision and integer related apps run great on Nvidia. Also, i don't know if you were around when PrimeGrid made their first GPU app, but it was Nvidia only, and remained that way for quite awhile. And it was great. But it also did not require double precision :)
EDIT: also, you said "If that is true". There is no "if" lol it is a fact. Though don't take my word for it. Here is an article from Anandtech posted just yesterday telling people about it http://www.anandtech.com/print/5699 scroll down just a little to the chart and view the FP64 column.
Also, here is forum thread from Nvidia's own forums from 2 years ago where a couple dozen people, including me, argued about it for a week or more http://forums.nvidia.com/index.php?showtopic=164417&st=20
By design, Nvidia's GPUs will do double precision at half the speed of single precision. Combine this with the intentional 75% penalty on the GTX 580, and it operates at 1/8 speed. For the GTX 680, it is the same, except the penalty is now 1/12th, resulting in 1/24th double precision speed, meaning the 580 is in fact faster than the 680 at double precision. Any benchmark will show you this to be a fact, such as this one: http://www.tomshardware.com/reviews/geforce-gtx-680-review-benchmark,3161-14.html However, the 680 is much faster than the 580 at single precision and integer calculations.
Alternatively, the AMD 7970 does double precision at 1/4th single precision speed by design, however it is not crippled, resulting in a 6x increase over the GTX 680.
|
5)
Message boards :
Number crunching :
Kepler rumors
(Message 51931)
Posted 4085 days ago by John P. Myers
Just so everyone is aware, Nvidia intentionally cripples the double precision performance of their graphics cards, which is why the BOINC community prefers OpenCL apps so we can use AMD GPUs, which are many many times faster.
In the case of the GTX 680, double precision is reduced intentionally, to 1/24th the speed of single precision. The GTX 580 was intentionally reduced to 1/8th.
Nvidia began doing this starting with the 400 series.
The 680 is a very good, very powerful graphics card. However, for us folders, if you intend to do any projects requiring double precision, Nvidia is NOT the GPU you want.
And likewise, if you're a project dev and will need to use double precision in your apps, CUDA is not the way to go. OpenCL is. An AMD 7970 is about 6x faster in double precision than a GTX 680.
|
6)
Message boards :
Number crunching :
Total number of WUs
(Message 36011)
Posted 4406 days ago by John P. Myers
Talking about the maximum which assumes settings are set to allow the maximum. But already got the answer. Thanks.
|
7)
Message boards :
Number crunching :
GTX 590
(Message 36008)
Posted 4406 days ago by John P. Myers
I have 2 EVGA GTX 590 Classified Hydro Coppers, both with a 14% overclock and run at 40C. I get about 2,400,000 credits/day on PG. I had it stable at a 24% overclock, but the memory on the 590's is on the back of the card instead of under the waterblock. They get hot enough to remove your fingerprints at that setting, even when leaving memory speed at stock, which i did.
|
8)
Message boards :
Number crunching :
Total number of WUs
(Message 36007)
Posted 4406 days ago by John P. Myers
Is there a limit on the total number of PG WUs you're allowed to have on your computer at any one time? Not talking about WUs that are part of a challenge. Just WUs in general. Thanks
|
9)
Message boards :
Number crunching :
Earth Day Challenge scoring discrepancy
(Message 22731)
Posted 4785 days ago by John P. Myers
Well...seems to be returning invalid work units hoping no one will notice so they can easily win. Afterall, they have been talking trash the last several weeks saying it would be nearly impossible to beat them in this race. I guess if you cheat, that's makes it true.
|
10)
Message boards :
Number crunching :
The Ides of March Challenge
(Message 21841)
Posted 4823 days ago by John P. Myers
15 minutes in and i can't get a single WU. this is rediculous.
|
Next 10 posts
|
