GPU Price Performance Curve

Robert
Robert
Joined: 5 Nov 05
Posts: 47
Credit: 318245623
RAC: 22155
Topic 195983

I put together this GPU Price Performance Curve while trying to decide on a recent GPU upgrade and thought it might be useful to others.

I averaged the elapsed times for a single BRP4 job from several different computers with a single GPU installed from the top computers list. For prices I checked newegg and picked the price of a card I would buy. I limited the candidates to GPU's in the Fermi series with good availability.

I avoided using the times from GPU's running multiple jobs simultaneously, they are usually obvious from the longer than average elapsed times. Running multiple jobs will of course increase the price / performance ratio of those cards, but it requires building an app_info file, which I personally don't want to deal with right now.

Two cards sit on the knee of the curve, the GTX 460 and the GTX 560 Ti. The 460 can be a real bargain when on special.

Power wise, the GT 430 requires no additional power connectors and can be very cheap, a good way to try out GPU computing before investing too much. Both the GT 450 and GTX 550 Ti only require a single extra 6 pin power connector, the rest of the cards require at least 2 six pin power connectors.

dskagcommunity
dskagcommunity
Joined: 16 Mar 11
Posts: 89
Credit: 1157322184
RAC: 253763

GPU Price Performance Curve

nice thx :)

DSKAG Austria Research Team: [LINK]http://www.research.dskag.at[/LINK]

archae86
archae86
Joined: 6 Dec 05
Posts: 3145
Credit: 7024634931
RAC: 1810885

Excellent. Too seldom do

Excellent. Too seldom do people think this way. Thanks for the posting.

What would make this much better, but would be much harder to obtain, would be data for power consumption vs. BOINC output.

One needs to consider purchase price and power cost separately, as the wide variation in the price of power moves the relative importance around for different users. But I suspect for many users for these cards the predicted lifetime cost of extra system power consumed (discounted properly to the purchase date if you want to be really careful) may be material compared to purchase price.

Sadly I doubt manufacturer's information would be much help. One would probably need data provided by users on the project of interest who happen to own suitable power meters (usually the ubiquitous Kill-a-Watt) or who are careful and clever enough to glean the required information from their residence meter.

archae86
archae86
Joined: 6 Dec 05
Posts: 3145
Credit: 7024634931
RAC: 1810885

As it happens I've been

As it happens I've been considering a new PC build for months. This would not be my primary daily use machine, but would be a full-capability machine with an extra role in hosting the hard drives which back up my other three PCs. I wanted to spend some extra for good BOINC performance without getting crazy expensive nor consuming a lot of power. So far this had me considering low end Z68 motherboards with a mid-upper Sandy Bridge CPU, probably 2500K. This thread jolted me out of my complacent plan to save power and cost by relying on the Sandy Bridge graphics. It appears that on a fleet-wide cost and power vs. BOINC performance basis it would make a lot of sense to build the system with a GTX 460 card.

If any folks here would care to comment on favorite GTX 460 cards who are using them for Einstein I'd love to hear it. My preliminary scans have me thinking about:

GIGABYTE GV-N460OC-1GI
EVGA 01G-P3-1371-AR

I'm especially interested in very low fan noise at idle, and moderate fan noise at full BOINC use. If you happen to have comparative experience on this point with more than one 460 card I'd be especially pleased to hear about it. I'm not much interested in burning a lot of extra power to get the last 10% of performance at the margin, and don't plan to overclock either the CPU or the GPU on this system. I'm also not interested in using a gamer-grade monster power supply, and would be please to hear if people successfully using this type of card for Einstein with good-quality supplies in the 430W range.

hotze33
hotze33
Joined: 10 Nov 04
Posts: 100
Credit: 368387400
RAC: 1355

Hi, I have the gigabyte card.

Hi,
I have the gigabyte card. While idle the fan is inaudible. The card is still relativly silent when under full load. I had a Seasonic 380W PSU powering this graphic card using the 6pin PCIe connector plus a molex to PCIe adapter. Under full load this card uses roughly 100W.

Robert
Robert
Joined: 5 Nov 05
Posts: 47
Credit: 318245623
RAC: 22155

Here are a couple more points

Here are a couple more points of reference on power to process BRP4 jobs (above the idle draw) for the cards I have used.

GT 430 = 26 watts
GTX 550 Ti = 73 watts

Bikeman (Heinz-Bernd Eggenstein)
Bikeman (Heinz-...
Moderator
Joined: 28 Aug 06
Posts: 3522
Credit: 686138814
RAC: 555564

Nice! What about also

Nice!

What about also plotting the recriprocal: units/sec vs price (or units/h vs $ for a more intuitive measure), and as a scatter plot, not a curve of linked points. This would give a nice view of productivity per $ (steeper slope if you link the point to the origin => better performance per $).

HBE

Stranger7777
Stranger7777
Joined: 17 Mar 05
Posts: 436
Credit: 417525414
RAC: 33714

RE: GT 430 = 26 watts GTX

Quote:
GT 430 = 26 watts
GTX 550 Ti = 73 watts


GTS 450 (Gigabyte with 1G) = 62 watts (1 WU), 72 watts (2 WU), 89 watts (3 WU), 51 watts (4 WU).
P.S. BOINC = 4 watts. ;)

Robert
Robert
Joined: 5 Nov 05
Posts: 47
Credit: 318245623
RAC: 22155

Here you go Bikeman - A GPU

Here you go Bikeman - A GPU Value Plot

This type of plot requested by Bikeman is a good place to show the value of a high end GPU running multiple jobs. So I looked at the current top computer with 2 x GTX 570 GPU's running multiple jobs (how many I can't say, maybe Aron can clue us in) and determined the number of BRP4 jobs crunched in a day. Specifically, I counted the number of jobs crunched in 3 days divided by 3 and divided again by 2 GPUs = 68 BRP4 jobs. 68 jobs * 500 credits = 34,000 credits / day, which is nicely ~half of the current 67,000 RAC of this machine.

The vertical axis is the number of BRP4 credits generated per day: (seconds in day / BRP4 elapsed time) * 500 credits. For GPU price I used the same value as in the price performance curve chart. As Bikeman points out, the steepest slope represents the best value, so as a visual aide I plotted the best value line. As you would expect from the price performance curve, the first 4 cheapest GPU's are all equal value.

You can see that the GTX 570 running multiple jobs has a better value than a GTX 570 running a single job, by the fact that it is closer to the best value line.

Bikeman (Heinz-Bernd Eggenstein)
Bikeman (Heinz-...
Moderator
Joined: 28 Aug 06
Posts: 3522
Credit: 686138814
RAC: 555564

Hi! Cooool!!! And it's

Hi!

Cooool!!! And it's nice also to see that the budget model in this field actually has a very good price/performance ratio.

Thanks for sharing this

HBE

joe areeda
joe areeda
Joined: 13 Dec 10
Posts: 285
Credit: 320378898
RAC: 0

Thanks to all who posted. I

Thanks to all who posted. I recently made this decision based on the "relative compute power" that nVidia publishes http://www.nvidia.com/object/graphics_cards_buy_now.html vs the price for the cards on Amazon.

Here is a price performance graph based on the manufacturer's published numbers. The image is big to see the labels. The data is pretty consistent with what's been posted already.

I picked the GTX 560 and run it on an I7-2600K machine with 16GB RAM (I have to recheck the speed, I think it's 1300MHz) it gets an average credit of about 20K. Other I7 machines using I've built average 6,000-7500 RAC so I put the GTX 560 at about 13,000 credits. Lower numbers than above but it's what I see.

I also have a few AMD Phenom II systems. One with a GT 240. Comparing that to other Phenom II's I estimate that one is worth about 5,000 RAC, it's not as clean because those systems are busier with real work.

Nothing is overclocked and I run one task at a time on the GPU.

Joe

Here's the table I used. Data is published relative compute (not display) performance and price is best I found on Amazon.com most had 1GB.

[/url]

Comment viewing options

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