Question re: Benchmarks Page

[Rodrigo]

Hello,

A quick question about the benchmarks page http://mersenne-aries.sili.net/bench.php:

The figures going across -- do they represent the performance of a single core from that CPU, or of the entire CPU? So for instance, if we compare the Intel Core i7 990X to the Intel Core i7-2600 (placed next to each other for your convenience , and for simplicity of comparison not even overclocked), would the numbers mean that as a whole the quad-core i7-2600 yields more output than the hex-core i7-990X... or that each of the 2600's four cores is more productive, individually, than each of the 990X's six cores? (Leaving aside the purchase price factor...)

Given the way the data is presented, I'm pretty sure that it's per core rather than for the whole CPU, but I want to verify this. I guess that, bottom line, what I'm looking for are straight-up Prime95 output comparisons of whole processors at factory settings.

Rodrigo

[James Heinrich]

It is for a single core. Data is taken from the Prime95 benchmark output, and for multi-threaded tests only the first (one thread running) result is recorded. No attempt is made to account for overall throughput of multi-core machines.

[Rodrigo]

Thanks, James -- much appreciated!

I can use the available data to apply some basic math (x4, x6) and get an estimate of the relative output. The head-to-head comparison feature comes in handy for this, as it gives the GHz-days per day.

Rodrigo

[henryzz]

Remember that there is some slowdown when running multiple cores. Core 2 quads are the worst at this(typing from a Q6600.) Running one of the cores tfing rather than lling was highly recommended on the first quads.

[Rodrigo]

Originally Posted by henryzz

Remember that there is some slowdown when running multiple cores. Core 2 quads are the worst at this(typing from a Q6600.) Running one of the cores tfing rather than lling was highly recommended on the first quads.

Huh, interesting. I didn't know that.

Would the same thing apply to the 980/990X, or to the new Sandy Bridge processors?

Or, would that also apply to this CPU: http://www.newegg.com/Product/Produc...ab=true&Page=3. (It goes in and out of stock.) Been thinking about putting that into my main PC to replace the factory-installed Intel Pentium Dual E2200.

Rodrigo

[Prime95]

Originally Posted by Rodrigo

Would the same thing apply to the 980/990X, or to the new Sandy Bridge processors?

No. Only the first generation of Core 2 processors (the Conroes?) had severe bandwidth problems.

You can expect a little bit of slowdown running 4 cores instead of 1 on later Core 2's. I've no experience with AMD processors. I expect they are more likely to experience bandwidth slowdowns running multiple workers.

When purchasing, keep in mind that Sandy Bridge will get faster in version 27. I've just begun writing FFTs using AVX instructions. It is a slow process, so beta is many, many months away.

[James Heinrich]

The Q6600 (I also have one) is horrible at multi-threaded throughput. Running 1, 2, 3, 4 threads on the same worker gives you something like (off-the-top-of-my-head numbers, I don't feel like actually benchmarking right now) 100%, 180%, 225%, 250% throughput of a single thread.

Even first-generation i7 (e.g. i7-920) is hugely better at scaling. The Q9xxx series (like the Q9550 you linked) is essentially part of the same design as the Q6600. While the Q9550 is undoubtedly much better than an E2200, I personally wouldn't spend anywhere close to the US$350 listed for the upgrade. I'd either put that towards a whole new SandyBridge machine, or if you want to do a small-price upgrade get a mid-range graphics card for half that price (e.g. GTX 460 for <$150) and use both cores of your E2200 to drive it.

[Rodrigo]

Originally Posted by Prime95

When purchasing, keep in mind that Sandy Bridge will get faster in version 27. I've just begun writing FFTs using AVX instructions. It is a slow process, so beta is many, many months away.

Cool!

Thanks much for the information. In light of what James says above, maybe the best bet will be to go (eventually) for a new PC with a Sandy Bridge processor.

BTW, is there any chance of getting graphics cards to run LLs with automated communication with the Prime95 server?

Rodrigo

[Rodrigo]

Originally Posted by James Heinrich

The Q6600 (I also have one) is horrible at multi-threaded throughput. Running 1, 2, 3, 4 threads on the same worker gives you something like (off-the-top-of-my-head numbers, I don't feel like actually benchmarking right now) 100%, 180%, 225%, 250% throughput of a single thread.

That's an enormous dropoff, whoa!

Originally Posted by James Heinrich

Even first-generation i7 (e.g. i7-920) is hugely better at scaling. The Q9xxx series (like the Q9550 you linked) is essentially part of the same design as the Q6600. While the Q9550 is undoubtedly much better than an E2200, I personally wouldn't spend anywhere close to the US$350 listed for the upgrade. I'd either put that towards a whole new SandyBridge machine, or if you want to do a small-price upgrade get a mid-range graphics card for half that price (e.g. GTX 460 for <$150) and use both cores of your E2200 to drive it.

Thank you for the details, this helps to guide my shopping. Sounds like the Q9550S would suffer the same kind of performance dropoff as the Q6600.

Reading the reviews for the GTX 460 card on Newegg, it looks like I'd also have to upgrade my 250W PSU to run it, unfortunately. I'd have to search for a card with more modest power requirements. Either that, or get a Sandy Bridge machine as you suggested.

Rodrigo

[James Heinrich]

Originally Posted by Rodrigo

looks like I'd also have to upgrade my 250W PSU to run it

Definitely. And any videocard low-end enough to actually work with a 250W PSU is unlikely to provide a worthwhile amount of performance.
Unfortunately it sounds like your system is sufficiently crippled (older CPU socket, small PSU) that any major performance-boosting upgrades don't make a lot of economic sense. Save the money and get a new Sandy Bridge system.

[Rodrigo]

Yes, it's looking more and more like that's the way to go.

Just one question about something you mentioned earlier:

Originally Posted by James Heinrich

I'd either put that towards a whole new SandyBridge machine, or if you want to do a small-price upgrade get a mid-range graphics card for half that price (e.g. GTX 460 for <$150) and use both cores of your E2200 to drive it.

Suppose that I were using a graphics card to crunch the numbers, and the two CPU cores to drive it -- how would that affect a PC's general computing performance (word processing, web surfing and the like)?

The reason I ask is that I've been thinking about doing GPU crunching, but this is my main computer where I do my work. If (like Prime95) it doesn't perceptibly affect performance, then I could run a graphics card for GIMPS either here or on a new Sandy Bridge system. But if it does noticeably affect responsiveness, then I should probably dedicate a machine to that setup.

Rodrigo