single core thru put of i7 920 versus i7 980x ...

[PsiStar]

With all of the benchmarks running multi-threaded, how does you make a single threaded comparison of these CPUs?

Hyper threading on.

Overclock the i7 920 to 3.33 GHz, same as the i7 980x (or is it)?

I have software that is both multi-threaded & single threaded depending on what is happening. Sometimes the single threaded processes can take an hour by themselves. So if the 980x is slower (single threaded) I am curious by what degree.

FWIW the multi-threaded ops can launch 16 threads.

 

[SHAQ]

Yep same clocks same motherboard. Force program to use 1 core and check the results. Some motherboards can force the CPU to use 1 core only via the BIOS.

 

[PsiStar]

I wasn't very clear, I do that a lot & didn't title the thread very well ... ok, my bad.

No. No forcing of a program to use 1 core as that is artificial. The software I am using has a variety of pre-processors, solvers, and post processors. In 1 run any combination of single threading and multi-threading processes occur ... sequentially.

So, with hyper-threading enabled as at least the main process uses all possible threads, is the why to, "how does the i7 980x compare to an over clocked i7 920 (to 3.33 GHz) for a single thread?"

I am considering swapping the i7 920, that I have, for a i7 980X in the ASUS P658D Premium m/b ... that I have. FWIW I don't screw around with OC-ing. I OC as much as possible!! The i7 920 is running at 4.41 GHz at present. I'll probably picket Intel (as rubycon once said) if I can't equal that ... well the goal is a i7 980X at 4.41 GHz in other words.

What I think I have been able to get out of a few benchmarks is that the i7 980X is not quite as efficient as the i7 920 for single threaded apps. I am hoping for some argument or a link or 3 to something to confirm, or not.

 

[aigomorla]

the 980X will scale higher then the 920.

Why? because u got more cache helping you with operations on the 980 then you do with the 920. (it maybe very small, but it should in application scale higher because of the 4meg extra cache.)

The 980X will also overclock better... (majority of the time) better then the 920, so u will have a higher ceiling in single core benchmarking.

If your going to do just single threaded marks, then a 980X would be financial suicide, because u will only use 1/6th of what the cpu can possibly do.

It doesnt make sense to run single thread benchmarks on a 980X, because u get the 980X for its 6c/12t awesomeness.... not because it can run single threaded applications super fast.

 

[Rubycon]

Single Core performance on Nehalem will beat Gainstown/Westmere IF it's highly dependent on cache latencies. For the most part the 32nm part will win clock for clock. It's also cooler running.

 

[VirtualLarry]

I seem to recall something about Intel making a compromise on the caches on the six-core. Bigger cache, but slower, or something like that.

 

[Rubycon]

I seem to recall something about Intel making a compromise on the caches on the six-core. Bigger cache, but slower, or something like that.

It is and this is why it makes little sense to go this route if you do nothing but game. Encoding / rendering OTOH makes all the difference! A stock 980X will walk all over a 920 o/c to 4.4 in this regard. I know because I have both and tried. When you overclock the 980 even to a (easy as taking candy from a baby) level of 4GHz, forget it! Contest over. Rendering performance is comparable to a dual socket 1356 Xeon (3-3.2GHz) system that costs WAY more.

Even though $1k for a chip is expensive it does provide a good value IF you put it to use.
The cooler Intel supplies with the chip is good enough to o/c to 4-4.2GHz so if you spent all your money on the chip and cannot afford a cooler you're set.

 

[PsiStar]

In this crowd, I feel like I am splitting hairs, but I am not talking about single core performance. Single threaded, yes, but on multi-core hyper-threading enabled CPUs with equal clock rates.

Not being grouchy ... just want to make sure that we are all on the same sheet of music.

For anyone's curiosity, the software works like this. You muck around & create a "job". It first meshes a 3D set of objects & is multi-threaded, but weakly in that only a few threads of what is available are used. Keep in mind that not all algorithms are equally efficient with 4 threads versus 8, just to pick some numbers. Those additional 4 threads may only buy a 10% improvement.

Then matrices are setup as well as several other checks, this is all single threaded & can take as long as another hour, for what I have watched so far, sometimes only seconds ... so this is very model dependent, just FYI.

Then the truly manly work is performed ... the solver runs. The solver is the product of many post grad hours of physics, engineering, & com sci. and is considered to be what makes it all worth while. Although, the solver is pretty stupid as it assumes that everything is copacetic & it can perform its task until a defined end is reached. It is iterative (the most threaded part) until convergence ... the virtual dust has settled to a point where the "end" is close enough. Without buying additional licensing options the solver has the capability of up to 16 threads.

The solver is why I would get the i7 980X. The other steps are a bit of a damper on the idea. In the back of my head there is the possiblity that the overall time could take longer on the 980X versus the 920!!! Yes, the solver would be a screamer, but the software in the preceding steps and the post processing steps is not as refined.

There is also the point that the 920 I am using is OC-ed to 4.41 GHz ... the signature describes the system. Whatif. What if the 980X will OC only to 4.2 GHz?! I *really* want to OC the 980X to 4.41 GHz! This will be in current the water cooled system in the signature. And, I have no problem pushing the chip with voltage. The justification is not so much about $$ but that I have a series of big projects in my queue, finally.:hmm:Time is $$ once again.

Anyway, I am suddenly at a point where I can afford the 980X. And, I cannot find the ... what ever ... benchmarks that gave me the idea that the single threaded (or fewer threads than 12) thru-put of a 980X is not as good as a 920. This is what this thread is all about. Maybe I don't have a life, need to get out more, and I just dreamed it?:\:\

 

[Rubycon]

Buy four systems and run side by side for testing in YOUR environment.

As far as wanting 4.4 what I've found is these chips WILL hit a wall right around 4.3. Beyond that requires substantially more VCORE. If you need more performance just go with a dual socket board and run a pair of X5680s at 4GHz. You also have the benefit of larger memory support. It's costlier however most REAL projects tend to have lofty budgets.

 

[PsiStar]

... most REAL projects tend to have lofty budgets.

There is that.

But, I am not designing the next 5 year out CPU. Perhaps the SSO of a breakout for a FPGA or the socket for that 5 year out CPU. So the budget is somewhat less that a million

 

[Shmee]

Single Core performance on Nehalem will beat Gainstown/Westmere IF it's highly dependent on cache latencies. For the most part the 32nm part will win clock for clock. It's also cooler running.

wont this be negated if someone OC's the gulftown/westmere's uncore? or is there more to it than the uncore speed?

 

[Rubycon]

wont this be negated if someone OC's the gulftown/westmere's uncore? or is there more to it than the uncore speed?

No because I overclock both.

Typically I see 3.6GHz uncore with bclk ~ 200. With the 980 one can reach 4GHz uncore with bclk below 150. I prefer bclk >200 though.

 

[Ben90]

wont this be negated if someone OC's the gulftown/westmere's uncore? or is there more to it than the uncore speed?

While overclocking the uncore will help, it will not alleviate the root problem.

Gulftown has a 48 cycle l3$ access time, Nehalem has a 42 cycle. From what Ive seen, most of the time programs prefer the 50% larger cache of Gulftown, but there are situations where the 14% higher latency hurts performance:

To the OP:
In most situations, Gulftown is faster clock/clock on single threaded applications due to the huge cache, but it is possible for it to be lesser due to the reasons above. In either case the difference wont be that large. I wouldn't let the single threaded performance be that big of a factor in deciding if you should upgrade to the 980x.

Also, if its possible to run multiple instances of the application at the same time, you could alleviate the single threaded woes by running multiple instances. Depending on how long the solver takes this could be a viable option, or a waste of your time setting up.

 

[tweakboy]

Even if not a multithread. A app will use all cores.

thx
gl

 

[Ben90]

Even if not a multithread. A app will use all cores.

thx
gl

I grabbed that quote first. No one else can take it!

 

[Makaveli]

lol ben can I buy it off you?