GPU Price Performance Curve

archae86
archae86
Joined: 6 Dec 05
Posts: 3,146
Credit: 7,097,684,931
RAC: 1,381,848

joe, thanks for your data. I

joe, thanks for your data. I tried computing a $ per unit performance measure for your table, and got the general curve shape as "stay away from the really cheap ones, and the expensive ones get gradually less competitive as you move up above $200". More specifically , the five champions on this particular measure, were the 240, the 9800, the 550 TI, the 460, and the 560, all costing very close to $8 for one unit of what nVidia calls performance (close enough that rounding to integer performance would add more noise than the apparent differences).

Of course, if one regards the whole host as overhead to producing the BOINC output of the graphics card, then one must add in the host cost. Even with Gary Roberts class efforts to make low-cost hosts, this would move the preferred card type up in performance (or push one to use multiple copies of one of the good cards toward the higher end of the best price/performance range).

I suspect including, as one should, an estimate of lifetime power consumption cost discounted to purchase date would little alter the preference order, with the very low and very high end cards again losing out even more to the "mighty middle". The very high end cards would suffer yet a bit more if one quite properly charged them for the price of more expensive system power supplies than otherwise would be required.

The curve below is based only on the table provided in this thread by joe areeda, and may suffer error by typo in my transcription. On this curve, low is good.

joe areeda
joe areeda
Joined: 13 Dec 10
Posts: 285
Credit: 320,378,898
RAC: 0

Nice presentation

Nice presentation Archae.

I put that table in as an image because I can never get bb code to display tables in a readable format. Sorry to make you transcribe it. I put a csv version at the end of this post.

The only thing I'd argue with in your post is the statement "if one regards the whole host as overhead to producing the BOINC output of the graphics card".

I like the E@H project but not enough to build systems solely to run it. So the question for me is more, live with the chip set display or add a usable GPU. So I see the 560 as 12,000 credits per day for a $185 investment. My normal use funds the rest of it and video encoding makes the GPU nice to have.

Oh and as far as power consumption, I have a UPS on that system that reports wattage. Running a BRP4 task adds about 100W (867 KwH/year or $150/year at LA DWP prices). Resulting in $34/million credits which is about twice as cost effective compared to the CPU tasks.

I made a mistake in the power supply calculation. The nVidia site recommends at least a 750W supply for a system with a 560 and adding up the maximum power draw on all the components came out to a bit over 500W so I put in a 750W.

Typical power draw at idle is about 60W, maxing out CPU is another 100W and the 560 is another 100W so the usual draw running 100% is about 260W. A 500W supply would have been plenty. So my numbers are somewhat higher than they should be.

Oh and my CPU and GPU temps are in the mid to high 50's C.

Joe

Here's that table as csv:

card","price","compute power"
"Gtx 590",$749.99,26
"Gtx 580",$489.99,31
"GTX 570",$344.99,30
"GTX 560TI",$249.99,26
"GTX 560",$189.99,23
"GTX 480",$462.03,30
"GTX 470",$499.00,27
"GTX 465",$279.00,23
"GTX 460",$159.99,20
"GTX 550 TI",$129.99,16
"GTX 450",$129.99,13
"9800 GT",$69.99,9
"GT 240",$70.00,9
"GT 440",$79.99,7
"GT 430",$66.32,6
"GT 520",$47.16,5
"9500 GT",$58.58,3
"9400 GT",$47.99,2
210,$34.99,2
"8400 gs",$31.56,1
MAGIC Quantum Mechanic
MAGIC Quantum M...
Joined: 18 Jan 05
Posts: 1,722
Credit: 1,106,815,887
RAC: 1,234,066

I will have to add some

I will have to add some nVidia cards myself since you have tempted me for quite a while now and I have 750 watt power supply's already.

This is where I always get my parts and machines since 1999 and they pretty much always have the best deals.

http://tinyurl.com/3zwy2vg

Sometimes you have to look around different pages since I did see the new GTS 8800's listed but they didn't have them yet.

http://tinyurl.com/GTS-8800

Of course those nVidia Tesla's are a bit expensive.

But the main thing is the ones that work here.

I got my power supplies there too.

The "kill-O-Watt" price doesn't bother me where I live.

(nice graphs btw)

tolafoph
tolafoph
Joined: 14 Sep 07
Posts: 122
Credit: 74,659,937
RAC: 0

RE: Sometimes you have to

Quote:
Sometimes you have to look around different pages since I did see the new GTS 8800's listed but they didn't have them yet.

What do you mean with "new GTS 8800". These are 5 years old and out of stock. http://en.wikipedia.org/wiki/GeForce_8_Series
I have one but upgraded to a GTX 260 almost 2 years ago.

Sascha

archae86
archae86
Joined: 6 Dec 05
Posts: 3,146
Credit: 7,097,684,931
RAC: 1,381,848

RE: I put that table in as

Quote:
I put that table in as an image because I can never get bb code to display tables in a readable format.

I used to use the code tag for this purpose, but pre tag is the one you want where the project supports it. I think this one does now--see example following

[pre]With the pre tag, multiple spaces, such as these display properly
and columns of data will align, such as these:
label value1 value2
first 1 1.001
second 2 2.002[/pre]

Quote:
Here's that table as csv:


Thanks. I've redone the graph using direct copy and paste from your csv. My eyes don't spot the difference, so the typo level may have been modest. I saved a new capture to Photobucket under the old name, so fresh displays of this thread will see the (slightly) revised version--including a second title line crediting you. A browser reload may be required to see the revised version for people who viewed the thread before my update.

MSE29
MSE29
Joined: 19 Aug 06
Posts: 14
Credit: 425,169,036
RAC: 88,200

Where in the ranking is a GTS

Where in the ranking is a GTS 450?
Or is it the same like GT450?

Much thanks in advance!

Best regards

Facebook | Twitter

Bikeman (Heinz-Bernd Eggenstein)
Bikeman (Heinz-...
Moderator
Joined: 28 Aug 06
Posts: 3,522
Credit: 692,136,310
RAC: 31,349

AFAIK there is no GT 450,

AFAIK there is no GT 450, it's called GTS 450 really.

CU
HBE

MAGIC Quantum Mechanic
MAGIC Quantum M...
Joined: 18 Jan 05
Posts: 1,722
Credit: 1,106,815,887
RAC: 1,234,066

RE: RE: Sometimes you

Quote:
Quote:
Sometimes you have to look around different pages since I did see the new GTS 8800's listed but they didn't have them yet.

What do you mean with "new GTS 8800". These are 5 years old and out of stock. http://en.wikipedia.org/wiki/GeForce_8_Series
I have one but upgraded to a GTX 260 almost 2 years ago.

Sascha

Yeah I didn't mean to say that.....see what happens when you do this at 4am

Which is why they don't sell those anymore.

archae86
archae86
Joined: 6 Dec 05
Posts: 3,146
Credit: 7,097,684,931
RAC: 1,381,848

I plan to place my order for

I plan to place my order for the new parts of a rebuild host by tomorrow, currently planning to use a Gigabyte 460 of the SOC variation, considering good reviews for that specific variant, and a very hefty rebate card for October purchases. I suspect that perhaps others will consider adding a CUDA card to a new build, or upgrading an existing system, or even building a dedicated BOINC host with CUDA stimulated by this thread.

At the risk of serious thread hijacking, I'd like to pose some of the questions a person considering placing an order or going through setup might have. I hope people with answers (or links to answers) to any will post them here, that this thread might serve as a fruitful launch pad for new CUDA user success.

1. What is required of a CUDA host on Einstein?
this thread in this forum gives a list, but the requirement of CUDA driver >= 3.2 (>= 260.00) raises the question: where does one get that?

2. If a new host with a CUDA-capable card installs BOINC and signs up to Einstein, will the CUDA capability automatically be used, or is more required? What?

3. Should a typical user expect that active CUDA processing will have visible impact on interactive use of the host? Does the answer depend on the number of stream processers? (It seems to be true for Radeon GPU use on SETI with the current Raistmer applications, low end cards suffer serious screen hesitation). Can good general advice be given as to which users would best be advised to employ the "Suspend GPU work while computer is in use? " option in Computing Preferences as reached from the Your Account page?

4. It is possible to configure a CUDA host to run more than one simultaneous task using ap_info. Where may the most up-to-date advice on doing that be found? Can any guidelines be suggested as to which CUDA cards will benefit, and at least roughly how much? (I hazard a guess this may depend on both the number of stream processers and the particular application, so this might be a rather volatile answer).

5. There appears currently to be a widely used production CUDA BRP application, called BRP4 in web posts but tagged as BPR3cuda32 in some places. Is there likely to be a Gravitational Wave CUDA ap, say within the next year?

6. Same question for Gamma Ray Pulsar?

7. Generalizing questions 5 and 6, does it seem likely that the utility of CUDA processing to the Einstein project, the quality of the development environment available, the quality of results obtained, and available project resources make it likely that a pretty steady supply of CUDA work will be available from Einstein for at least the next couple of years?

8. Generally both the manufacturer's advice and web comments from game-oriented users seem to advise hyperconservative (excessive) power supply ratings for use with these cards. Any additional comments to those already in this thread (thanks, guys) on real world success with particular configurations and supply ratings would help us choose cards and configure hosts.

9. The schedulers seem to behave especially poorly in the presence of mixed applications and application types. Is there any rule-of-thumb advice that might help, especially a beginner? I'll offer one--start small (say four hours of requested queue time, or less). But in the current climate would a person using ap_info anyway to get >1 active CUDA task on one card be well advised to use the ap_info to segregate CUDA work and CPU work to completely different types of work? Or yet more strenuously, to devote GPU on one Einstein application and CPU to work from another project altogether?

10. All the other questions I should have asked but failed to which would help people to decide whether to add CUDA, to buy the right stuff, and successfully to start up their new capability.

Richard Haselgrove
Richard Haselgrove
Joined: 10 Dec 05
Posts: 2,142
Credit: 2,792,655,020
RAC: 695,795

RE: I plan to place my

Quote:

I plan to place my order for the new parts of a rebuild host by tomorrow, currently planning to use a Gigabyte 460 of the SOC variation, considering good reviews for that specific variant, and a very hefty rebate card for October purchases. I suspect that perhaps others will consider adding a CUDA card to a new build, or upgrading an existing system, or even building a dedicated BOINC host with CUDA stimulated by this thread.

At the risk of serious thread hijacking, I'd like to pose some of the questions a person considering placing an order or going through setup might have. I hope people with answers (or links to answers) to any will post them here, that this thread might serve as a fruitful launch pad for new CUDA user success.


0a. In all that follows, I assume you'll be running Windows?
0b. I can only speak about NVidia cards, not ATI

Quote:
1. What is required of a CUDA host on Einstein?
this thread in this forum gives a list, but the requirement of CUDA driver >= 3.2 (>= 260.00) raises the question: where does one get that?


For the latest (non-beta) driver, http://www.nvidia.com/Download/index.aspx?lang=en-us
For a full choice, http://www.nvidia.com/Download/Find.aspx?lang=en-us

Quote:
2. If a new host with a CUDA-capable card installs BOINC and signs up to Einstein, will the CUDA capability automatically be used, or is more required? What?


A decent, recent, BOINC. I'm happy with v6.12.34 - onther people prefer v6.10.58/60
Do not install in 'service' or 'PAE' mode.
Check all usage defaults are suitable, but from there on, it's automatic.

Quote:
3. Should a typical user expect that active CUDA processing will have visible impact on interactive use of the host? Does the answer depend on the number of stream processers? (It seems to be true for Radeon GPU use on SETI with the current Raistmer applications, low end cards suffer serious screen hesitation). Can good general advice be given as to which users would best be advised to employ the "Suspend GPU work while computer is in use?" option in Computing Preferences as reached from the Your Account page?


With decent kit, I've never found it to be a problem. I'd start with the presumption that you do not need to suspend GPU work for daily computer use. Exceptions might be gaming and DVD/TV playback, but I haven't tried any of those.

Quote:
4. It is possible to configure a CUDA host to run more than one simultaneous task using ap_info. Where may the most up-to-date advice on doing that be found? Can any guidelines be suggested as to which CUDA cards will benefit, and at least roughly how much? (I hazard a guess this may depend on both the number of stream processers and the particular application, so this might be a rather volatile answer).


It'll only be worth trying this on 'Fermi' class cards - 4xx or 5xx series. (they have hardware acceleration for context switching). 'How many' is dependant on GPU RAM and the size (memory footprint) of the tasks being run.

'Multiple tasks per card' and 'multiple projects' can lead to scheduling problems. If you run two Einstein tasks, and one of them finishes, BOINC has to find a 'half GPU' task to fill the vacancy. If your other project requires a complete card, it never runs....

Quote:
5. There appears currently to be a widely used production CUDA BRP application, called BRP4 in web posts but tagged as BPR3cuda32 in some places. Is there likely to be a Gravitational Wave CUDA ap, say within the next year?


Pass

Quote:
6. Same question for Gamma Ray Pulsar?


Pass

Quote:
7. Generalizing questions 5 and 6, does it seem likely that the utility of CUDA processing to the Einstein project, the quality of the development environment available, the quality of results obtained, and available project resources make it likely that a pretty steady supply of CUDA work will be available from Einstein for at least the next couple of years?


Pass

Quote:
8. Generally both the manufacturer's advice and web comments from game-oriented users seem to advise hyperconservative (excessive) power supply ratings for use with these cards. Any additional comments to those already in this thread (thanks, guys) on real world success with particular configurations and supply ratings would help us choose cards and configure hosts.


Check specifically the 12v rating on the PSU. I've been advised to look for 'single rail' 12v designs. Some cards - especially the manufacturer overclocked 560Ti - have been noted as being vulnerable to producing error reults if the power supply is inadequate (and some users have countered that tendency by over-volting using OC tools)

Quote:
9. The schedulers seem to behave especially poorly in the presence of mixed applications and application types. Is there any rule-of-thumb advice that might help, especially a beginner? I'll offer one--start small (say four hours of requested queue time, or less). But in the current climate would a person using ap_info anyway to get >1 active CUDA task on one card be well advised to use the ap_info to segregate CUDA work and CPU work to completely different types of work? Or yet more strenuously, to devote GPU on one Einstein application and CPU to work from another project altogether?


Einstein has a handy web preference for "Run CPU versions of applications for which GPU versions are available". I have that set to 'no', but apart from that (and the fractional GPU usage noted above), I've had no scheduling problems.

There are concerns at the moment that GPUs may lock into a protective downclocked mode if applications are not written to flush and close all threads safely before exit - the faster the card, the more likely it is to be vulnerable to this. Einstein's current app is not yet threadsafe and may trigger this. I've never suffered it myself, but to be on the safe side, I've increased BOINC's 'task switch interval' to longer than the observed BRP4 runtime, so that tasks complete in one run and aren't pre-empted part-way through.

Quote:
10. All the other questions I should have asked but failed to which would help people to decide whether to add CUDA, to buy the right stuff, and successfully to start up their new capability.


It's all actually very straightforward these days. After general experience with BOINC, I configured a CUDA-enable laptop (host 3868392) with the minimum of fuss this year. The hardest part was convincing Windows 7 that I didn't want to save power, but wanted it to run at full speed, with cooling to match.....

Enjoy.

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.