• Welcome to Overclockers Forums! Join us to reply in threads, receive reduced ads, and to customize your site experience!

How many FLOPS?? (and which CPU)

Overclockers is supported by our readers. When you click a link to make a purchase, we may earn a commission. Learn More.

4GHZ_or_bust

Now 6GHz or Bust!
Joined
Aug 1, 2002
Location
Michigan
Was looking for a chart giving an average FLOPS per CPU, and found this one on Wiki:

* Folding@Home, as of March 11, 2010, is sustaining 3.874 PFLOPS [1], the first computing project of any kind to cross the 1,2,3,4 and 5 petaFLOPS milestone. This level of performance is primarily enabled by the cumulative effort of a vast array of PlayStation 3 and powerful GPU units.[2]
* The entire BOINC project averages over 4.725 PFLOPS as of Mar 11th 2010[3].
* Milkyway@Home averages over 1.551 PFLOPS as of Mar 11th 2010[4]. A press release reported that MilkyWay@Home surpassed the 1 Peta Flop barrier (1.009PFlops) on Jan 26th 2010.[5]
* SETI@Home computes data at an average of more than 827 TFLOPS[6] as of Mar 4th 2010
* Einstein@Home is crunching more than 238 TFLOPS[7] as of Nov 1st, 2009
* As of Nov, 20th 2009[update], GIMPS is sustaining 44.3 TFLOPS.[8]

Just... wow :shock: that's a lot of FLOPS to flip.

I did eventually find a chart of FLOPS by CPU:
10,222.58 MFlops Intel Core i5 750 (4105MHz)
9,252.76 MFlops Intel Core i5 750 (3675MHz)
7,841.57 MFlops Intel Core i5 750 (3161MHz)
7,525.77 MFlops AMD Phenom II X4 940 (4258MHz)

It seems to me that P2 X4 overclocked to 4.25GHz is doing close to i5 750 clocked to 3.2GHz. (I couldn't find Athlon X4 FLOPS rating) And since i5 750 can easily overclock to 4GHz+, it almost looks like it'd be worth tossing an extra $100 to get i5 750 instead of AMD X4 for my second PC. The same chart showing i7 920 clocked to 4GHz has about slightly more FLOPS as i5 750 so it looks like BOINC does not benefit from hyperhtreading, only from tri-channel RAM.

Does that sound about right?
 
If you're putting together a system just for SETI I'd consider adding video cards to your current rigs instead of building a new one. A cheaper chip and motherboard with high-end video may also be a better option over the more expensive main system. The idea needs some research, as prices have changed since I looked over the options, but I do think it's worth a look ...
 
I wanted to build a second PC as backup to replace the vintage s478 P4 I still have around and it'd only make sense that it'd be able to do SETI@home while sitting around waiting for my main PC to suffer major failure.
 
I am really interested in the $$ per seti WU question. I have recently returned to seti, and have 2 quad cores and a dual core intels working on it, no GPUs yet, when I left, there weren't GPU clients .....

I have a couple amdx2 939 computers ..... I could drop GPUs in those ... or upgrade processors or mb + proc + ram or ... or or or :)

If I wanted to build just a cheap motherboard + cpu + ram combo for seti, what would be best?

In the past, I ran a network booted LTSP 12 headless motherboard yata processing farm .... (for the old folks that remember those :) Though it looks like just booting linux using cheap usbkeys might be about as cheap and easier.

IS THERE ANYWHERE were we can see the seti performance of various processors and / or GPUs? How does say an AMDX2 compare to a nvidia gforce 8800GTX?

What's the cheapest $$ per seti WU set up I can get, OR, where is the some data I can use on seti performance so I can calc it out?

Thanks much for any help!

I know quad cores with tri GPU 280X produce massive WU results, and are WAY sexy :) but in the past, cheap no name no overclocking (yep) motherboards with last years processor and NO video card were the cheap way to go. Multi cores and multi GPUs may be a cheaper way now, as you can reduce the number of powersupplys.....

A scondary cost consideration, if anyone has any experience, is in running power costs. Previous seti farms I've run could up the eletric bill over $100 per month. I wonder if a single 1000 watt processor with quad core cpu + 2 GPUs would consume more or less power than a setup of 350 watt powersupplies powering just a CPU each that would produce similar WU output? In other words, for a given number of WU, would beefier machines consume less total power than multiple lower CPU machines required to do the same amount of WU?

Thanks for any insights!
 
Was looking for a chart giving an average FLOPS per CPU, and found this one on Wiki:



Just... wow :shock: that's a lot of FLOPS to flip.

I did eventually find a chart of FLOPS by CPU:
10,222.58 MFlops Intel Core i5 750 (4105MHz)
9,252.76 MFlops Intel Core i5 750 (3675MHz)
7,841.57 MFlops Intel Core i5 750 (3161MHz)
7,525.77 MFlops AMD Phenom II X4 940 (4258MHz)

It seems to me that P2 X4 overclocked to 4.25GHz is doing close to i5 750 clocked to 3.2GHz. (I couldn't find Athlon X4 FLOPS rating) And since i5 750 can easily overclock to 4GHz+, it almost looks like it'd be worth tossing an extra $100 to get i5 750 instead of AMD X4 for my second PC. The same chart showing i7 920 clocked to 4GHz has about slightly more FLOPS as i5 750 so it looks like BOINC does not benefit from hyperhtreading, only from tri-channel RAM.

Does that sound about right?
Finally had time to do a little checking on that approach.
Phenom 940BE: 3.52GHz, 6.7 GFlops, 42-4800 RAC
C2Quad Q6600: 2.40GHz, 4.1 GFlops, 36-4200 RAC

Both of those rigs are dedicated SETI units and don't run CUDA. Obviously there's no clear correlation between GFlops and RAC (with the possible exception of CPUs in the same family of processor). The wide range of RAC is what I've seen over the past year and seems to vary as the average size of WU varies (though I can't be sure of that since it would be extremely time consuming to keep track). Each architecture works a little differently and even those two above vary in their differences by WU size with the AMD being more effecient on larger WU (maybe because of the L3 size? but maybe not :shrug: ).

I am really interested in the $$ per seti WU question.
I started a thread on this topic in '07 which lasted a few months and got some good results. If you'd like to reopen that thread or, better yet, start a new one I'm willing to discuss it ... :)
 
Last edited:
Performance for the $$ is Always going to favor the GPU's over CPU's, there's really no comparison. A GTX260 could produce as many credits as an overclocked i7. You don't need very powerful CPU's to support the multiple GPU's either. I can't really comment on the electricity costs beyond saying that a multi-GPU setup will be the most efficient.

As you can see by my sig, I'm a little biased towards that strategy, having filled my stable with Multi-GPU motherboards.
 
Performance for the $$ is Always going to favor the GPU's over CPU's, there's really no comparison. A GTX260 could produce as many credits as an overclocked i7. You don't need very powerful CPU's to support the multiple GPU's either. I can't really comment on the electricity costs beyond saying that a multi-GPU setup will be the most efficient.

As you can see by my sig, I'm a little biased towards that strategy, having filled my stable with Multi-GPU motherboards.
Aren't you CUDA guys getting slammed from lack of work right now or was that something else I was reading?


I DO agree with your basic assessment, though. GPUs are always more efficient and more economical than CPUs ...
 
in your question did you say that quads are to costly and your limited to a dual?
cause that's not the case Amd's got some awsome cheap quads that can be overclocked

Quote
"I DO agree with your basic assessment, though. GPUs are always more efficient and more economical than CPUs ..."

Gpu's are the highest power consumers in your computer
 
Here are some (theoretical) RAC numbers based on recent WU's. All are running optimized clients. Most of these are averages of at least 5 different sized WUs. These are very rough estimates; real-world numbers will certainly be lower. But the relative comparisons are useful.

Core i7 [email protected]...6500
Core 2 [email protected]....5000
GTX 295...25000
GTX 275...16000
GTX 260...13000
9800 GTX...9700
GTX 250...9000

Based on these numbers, you can compute RAC per dollar invested (higher is better) based on current NewEgg prices. From best to worst...

9800 GTX: 72.02
GTX 275: 62.96
GTX 260: 62.80
GTX 250: 57.10
GTX 295: 53.00
C2Quad: 29.67
Core i7: 22.30

Of course, keep in mind that the 295 has 2 GPUs per card, so you could in theory get 6 working in a single box, for a whopping ~75000 RAC just from the GPUs (minus overhead, of course).

As you can see, the CPU's are dismal especially when you consider that you need a motherboard, power supply, RAM etc. to run just one, but you can (in theory) run up to 3 of the GPU cards in a system.
 
Here are some (theoretical) RAC numbers based on recent WU's. All are running optimized clients. Most of these are averages of at least 5 different sized WUs. These are very rough estimates; real-world numbers will certainly be lower. But the relative comparisons are useful.

Core i7 [email protected]...6500
Core 2 [email protected]....5000
GTX 295...25000
GTX 275...16000
GTX 260...13000
9800 GTX...9700
GTX 250...9000

Based on these numbers, you can compute RAC per dollar invested (higher is better) based on current NewEgg prices. From best to worst...

9800 GTX: 72.02
GTX 275: 62.96
GTX 260: 62.80
GTX 250: 57.10
GTX 295: 53.00
C2Quad: 29.67
Core i7: 22.30

Of course, keep in mind that the 295 has 2 GPUs per card, so you could in theory get 6 working in a single box, for a whopping ~75000 RAC just from the GPUs (minus overhead, of course).

As you can see, the CPU's are dismal especially when you consider that you need a motherboard, power supply, RAM etc. to run just one, but you can (in theory) run up to 3 of the GPU cards in a system.

I'd say those GPU numbers are a little optimistic, and you can actually run 4 GPU's in one system. My K9A2 platinum supports 4 dual slot cards, my M3A32 MVP dlx supports 3 dual slot and 1 single slot cards.
 
For some reported GPU RAC's see post #1:
http://www.overclockers.com/forums/showthread.php?t=617684

Quote
"I DO agree with your basic assessment, though. GPUs are always more efficient and more economical than CPUs ..."

Gpu's are the highest power consumers in your computer
According to APC, and I admit that's a rough reading at best, most of my systems run ~200W under SETI load (or just a hair more). From what I've read isn't that about the same as the most power hungry GPUs? If so, the RAC/W on most GPUs is higher than running a full system, which is what you have to do to run one CPU.

Of course, I'm not running any i-series quads, either, so I can't say what those system eat for power but I'm sure once you take the total system cost into account as well as, say, three years of power usage you'll still end up with a higher cost per RAC for a complete system over buying and running another GPU or two in an existing system ...
 
Aren't you CUDA guys getting slammed from lack of work right now or was that something else I was reading?

There's plenty of work right now, but in general, CUDA rigs will run out of work faster than dedicated CPU rigs. This is mostly because It's impossible to set a high cache value on fast/multi-GPU CUDA boxes because the shear volume of tasks will cause BOINC to freeze. CUDA boxes will also clear the cached workunits faster. When there's a a SETI outage, we have smaller caches that we work through faster so it's more likely we'll run out of work and sit idle for a while. My RAC went into the 20k's during the most recent outage, but it's back up to almost 45k now.
 
I'd say those GPU numbers are a little optimistic, and you can actually run 4 GPU's in one system. My K9A2 platinum supports 4 dual slot cards, my M3A32 MVP dlx supports 3 dual slot and 1 single slot cards.

Yup, I realize they're optimistic. Basically all I did was take the average "run time" and calculate what it would produce in a day. I would estimate 10-20% overhead (minimum) which doesn't take into account other issues that may slow things down. Murphy has a way of ruining all theoretical numbers ;)

That's cool that there are 4 slot mobo's out there--but I thought I'd read that there is a BIOS or OS limit at 3. Am I wrong on that?
 
Maybe for Intels you're limited, I don't know, but AMD's have been 4-slot capable at 4x8 for well over a year now (790FX chipset). They don't run SLI for video purposes (Crossfire only) but they can support 4 nVidia boards for crunching, IIRC ...
 
Back