Do # of cores or GHz/per cores perform better?

The most demanding Projects are only available to a small percentage of very high-end servers.

Moderators: Site Moderators, PandeGroup

Do # of cores or GHz/per cores perform better?

Postby PinHead » Sun Feb 05, 2012 2:26 am

Recently back to folding, seems during a time of transition, and after days of scrolling and guessing search words I haven't been able to find the answer to the above question.

All things being equal, and by things I mean CPU's, which will perform better at bigadv for folding:

16 cores at 3GHz for a total of 48GHz?

or

32 cores at 1.5GHz for a total of 48GHz?


This is not a CPU architecture question, but a FAH client architecture question; with all things being equal which will perform better with the client?
PinHead
 
Posts: 351
Joined: Tue Jan 24, 2012 3:43 am

Re: Do # of cores or GHz/per cores perform better?

Postby Jesse_V » Sun Feb 05, 2012 2:40 am

Hmm. That's a very interesting question. I'm not sure I know enough to give you a proper answer! I mean I do know that there's a lot more to CPU speed than just the clock rate, but if like you say all things being equal then I would think that the 32 cores would be better. I mean I don't know all the intricacies F@h WUs (specifically bigadv) but I'd go with 32 because its my understanding that the protein being simulated is broken into "boxes" (a subset of the amino acids, like groups/neighbors of parts of the protein) where each processor takes a particular box and crunches it. Taking this analogy one step further, is a theoretical 1 core CPU at 48 GHz better/worse than a 32 core 1.5 GHz CPU? Parallel processing just seems to be to be the superior option: you can get more done at once, and since I'd suppose that the box for a 32 core CPU is half the size of a 16 core, there's less work for it to do in that box, thus the 32 core is better even at a reduced clock rate. I mean I don't know if a protein can scale like that. Like I said, that's a very fascinating question, and I'll have to go back and review some material I have on this subject, and I look forward to the expert responses. But I thought I'd through out my guess for what it's worth. Maybe it makes no difference. Get one of both and tell us the PPD! :D
Pen tester at Cigital/Synopsys
User avatar
Jesse_V
 
Posts: 2773
Joined: Mon Jul 18, 2011 4:44 am
Location: USA

Re: Do # of cores or GHz/per cores perform better?

Postby k1wi » Sun Feb 05, 2012 3:19 am

At best they would be equal, at worst there would be a small favor towards faster cores, assuming the architecture is the same. Folding@home tends to scale pretty well as you increase cores (though this is only known up to a particular point), so there is not too much cost to increasing core counts.

Unfortunately, it does become a CPU architecture issue, because there is a lot of sub-systems that can have a major effect on folding performance, such as the bus speed - which will tend to be at the same clock regardless of the number of cores being used - or even memory.

The easiest way to test it and come up with some solid results is to downclock a chip for one test and then disable a number of cores.
k1wi
 
Posts: 1152
Joined: Tue Sep 22, 2009 10:48 pm

Re: Do # of cores or GHz/per cores perform better?

Postby 7im » Sun Feb 05, 2012 3:27 am

I don't expect the speed between those two examples to be that much different. In the past, a 4 GHz chip earned twice as many points as the same chip running at 2 GHz. That P4 and single core CPU testing was done a long time ago. I don't recall where anyone has done this kind of testing lately with SMP, or SMP with QRB.

Assume both are the same speed. The 16 core box is cheaper to build, so then build more of them. 16 core wins. ;)
How to provide enough information to get helpful support
Tell me and I forget. Teach me and I remember. Involve me and I learn.
User avatar
7im
 
Posts: 14648
Joined: Thu Nov 29, 2007 4:30 pm
Location: Arizona

Re: Do # of cores or GHz/per cores perform better?

Postby PinHead » Sun Feb 05, 2012 4:35 am

7im wrote:I don't expect the speed between those two examples to be that much different. In the past, a 4 GHz chip earned twice as many points as the same chip running at 2 GHz.


I'm not sure that is a fair example since it is a one to one relationship and cutting a clock in half would yield the expected 1/2 results.

This is more like comparing 1 4 GHz chip to 1 Quad core 1GHz chip to determine if the Quad core 1GHz carries a 10% overhead or more. Then multiply it by 4 ( or more ) to scale to 16, 32, etc.

Now I'm wondering where the point of diminshing return happens to occurs whether it be hardware base or client base.
PinHead
 
Posts: 351
Joined: Tue Jan 24, 2012 3:43 am

Re: Do # of cores or GHz/per cores perform better?

Postby 7im » Sun Feb 05, 2012 5:41 am

My point is that ghz performance scales up linearly.and since you also scaled the core count in your example in a linear manner. They both come out about the same.

And from past postings, the client scales well up to 32 cores, up to 64 as I recall. It's not 100% but it is linear at least that far.

A person with 128 posted not to long ago, but I don't recall the info...

So, just curious, or are you looking to pull the trigger on some heavy metal?
User avatar
7im
 
Posts: 14648
Joined: Thu Nov 29, 2007 4:30 pm
Location: Arizona

Re: Do # of cores or GHz/per cores perform better?

Postby Nathan_P » Sun Feb 05, 2012 10:06 am

7im wrote:Assume both are the same speed. The 16 core box is cheaper to build, so then build more of them. 16 core wins. ;)


Not neccesarily, At the lower end of the intel CPU line up you can, however my dual xeon x5670 box cost about £1,500 to build - you can build a quad 8core opteron rig for the same or maybe even slightly less
Image
Nathan_P
 
Posts: 1442
Joined: Wed Apr 01, 2009 9:22 pm
Location: Jersey, Channel islands

Re: Do # of cores or GHz/per cores perform better?

Postby bruce » Sun Feb 05, 2012 8:45 pm

It also depends on what you call a "core" Both Intel and AMD have started packaging a pair of integer processing units with a single floating point processing unit. but then they're faced with a question of what to call them. (They've each got a sales department that establishes what they want to advertise.) Since FAH uses mostly floating point processing, a "core" that shares a FPU with another "core" is worth very slightly more than half of what a "core" that has it's own FPU is worth.
bruce
 
Posts: 22623
Joined: Thu Nov 29, 2007 10:13 pm
Location: So. Cal.

Re: Do # of cores or GHz/per cores perform better?

Postby Nathan_P » Sun Feb 05, 2012 8:55 pm

bruce wrote:It also depends on what you call a "core" Both Intel and AMD have started packaging a pair of integer processing units with a single floating point processing unit. but then they're faced with a question of what to call them. (They've each got a sales department that establishes what they want to advertise.) Since FAH uses mostly floating point processing, a "core" that shares a FPU with another "core" is worth very slightly more than half of what a "core" that has it's own FPU is worth.


But that gets even more complicated when AMD say that their single FPU can process 2 128bit instructions at the same time, this in effect turns the cpu back to an 8 core device - however there are only one of various other bits of the cpu per 2 core module.
Nathan_P
 
Posts: 1442
Joined: Wed Apr 01, 2009 9:22 pm
Location: Jersey, Channel islands

Re: Do # of cores or GHz/per cores perform better?

Postby bruce » Sun Feb 05, 2012 9:00 pm

They both come out about the same.

The real answer to the question is 1 core at 48GHz ;) but since nobody has figured out how to do that, the answer "about the same" it a good one. The difference is in the OS, not so much as in the hardware. The task scheduler knows exactly how to schedule work for a single thread. The OS has more trouble scheduling a task that has multiple threads. I'm sure you've seen topics that talk about how SMP slows down when other processes usurp some resources it would like to have, and the answer boils down to something like: SMP runs at the speed of the slowest thread. On a machine that runs nothing but FAH, that's not an issue, but you'd have to ask yourself whether in the actual environment you're going to be running, would the "typical" differences between the slowest and fastest thread be any different depending on whether you had 16 threads or 32 threads. I'm not even sure how to predict an answer to that question.
bruce
 
Posts: 22623
Joined: Thu Nov 29, 2007 10:13 pm
Location: So. Cal.

Re: Do # of cores or GHz/per cores perform better?

Postby bruce » Sun Feb 05, 2012 9:20 pm

Nathan_P wrote:But that gets even more complicated when AMD say that their single FPU can process 2 128bit instructions at the same time, this in effect turns the cpu back to an 8 core device - however there are only one of various other bits of the cpu per 2 core module.
Yes, it does get more complicated. Intel has also increased the effectiveness of their shared FPU with each major hardware redesign, but I'm not going to get into an argument over which is better or why.

In fact, it's not really clear how that advertising claim relates to FAH. The assembly optimizations use SSE which combines four single-precision operations into a single operation on any hardware that supports SSE (almost everything, nowadays). I often see advertised figures for double-precision GFLOPS or for double-precision SSE2 but gromacs doesn't use either one.

All advertisers are going to say things that make their product look better than the completion and it's up to the reader to know how that applies to them. The real data may be available in the hardware forum where folks report direct comparisons of FAH on two platforms of interest.
bruce
 
Posts: 22623
Joined: Thu Nov 29, 2007 10:13 pm
Location: So. Cal.

Re: Do # of cores or GHz/per cores perform better?

Postby PinHead » Mon Feb 06, 2012 1:36 am

7im wrote:So, just curious, or are you looking to pull the trigger on some heavy metal?


Well, I kinda pulled the trigger before I thought to ask that question. :eo

But the more I evaluated prices on 1 gen back 8 core cpu's at higher clock rates vs newer 16 core cpu's at lower clock rates; I noticed they were being priced on overall GHz ( cores x clock ). Pretty sure the manufactures are being controlled by a heat equation here.

So I bought a 16 with a motherboard that will let me drop another one in later. Say after a couple paychecks and checking in with Uncle Sam.
PinHead
 
Posts: 351
Joined: Tue Jan 24, 2012 3:43 am

Re: Do # of cores or GHz/per cores perform better?

Postby Slash_2CPU » Tue Feb 07, 2012 5:11 am

It would depend much more upon the architecture of those 16 or 32 cores. Those core counts imply that you are looking at AMD. Read up on the threads regarding Bulldozer-based CPU's(4200 and 6200-series Opterons) and PPD vs the last generation of K10 CPU's(4100/6100-series Opterons). The K10 is currently a better bang for the buck.
ASRock X99 WS i7-5930K @ 4.4GHz /2x GTX 970 @ 1.46GHz /4x4GB DDR4-2666
Phenom II X6 @ 3.7GHz /2x2GB DDR3-1680 /GTX 970 @ 1.40GHz
450-600K PPD @ ~850W
Slash_2CPU
 
Posts: 196
Joined: Sat Apr 19, 2008 5:15 pm

Re: Do # of cores or GHz/per cores perform better?

Postby Ivoshiee » Tue Feb 07, 2012 7:35 pm

Number of cores will triumph core speed if the selected project does scale. I have a feeling that current crop of projects does that.
Ivoshiee
Site Moderator
 
Posts: 1286
Joined: Sun Dec 02, 2007 12:05 am
Location: Estonia

Re: Do # of cores or GHz/per cores perform better?

Postby Nathan_P » Tue Feb 07, 2012 7:41 pm

Given the recent annonucements, Number of cores (min 16 for bigadv) followed very closely by the fastest clock speed for your chosen chips that you can afford.
Nathan_P
 
Posts: 1442
Joined: Wed Apr 01, 2009 9:22 pm
Location: Jersey, Channel islands

Next

Return to SMP with bigadv

Who is online

Users browsing this forum: No registered users and 1 guest

cron