Clarksfield v Sandy Bridge

Hi all,
An interesting observation on my two machines(see my specs in sig)

I began my notebook adventure with my clevo p170hm and it came with sandy bridge i7-2820QM. I also noticed that there was some stuttering occasionally in gaming back when I bought it and many said disable core parking and I did and no more stuttering and games were fluid afterwards
Then, I recently bought the Alienware which came with the Clarksfield i7-840QM. I immediately thought disable core parking too to get better gaming performance but here is what I strangely found.

Upon doing this, with Clarksfield, I did not get any significant improvement actually it felt worst in gaming. So I then looked at the core clocks(max, min, average) with hwinfo64 and noticed something intriguing. With core parking disabled(Clarksfield) the single peak freq on a single core was less than with core parking enabled. I then enabled core parking with the clarksfield and would you believe it the max peak freq hit about 3164.5Mhz on one of the cores which is approaching the max 3200Mhz. With core parking disabled I would only hit 3048Mhz across two cores peak using cpu-burn and this I felt in gaming too with Crysis 3 you can feel that the single core peak freq is preferable than the disabled core parking which lowers freq across all 4 cores.

So does this mean clarksfield has better single core performance? I recall many say the core parking is way too aggressive from windows but strangely the clarksfield feels up to it. I have yet to go and do this same test on the Sandy bridge and see what it's single core performance is like but I seem to think it could be worst off hence what caused the stuttering. Could one infer that Clarksfield has better single core performance than Sandy bridge?

Thought/opinions?

 

I don't understand. You saw a decrease in performance after disabling core parking on the 840qm. Then you saw a higher frequency when you enabled core parking. Why do you think that translates as better single threaded performance on the Clarksdale architecture? The freq. with parking disabled was lower probably because the system was designed to throttle or the older hardware wasn't able to supply as much current to increase frequency close to max across all cores. You can check this by checking the voltage for the CPU with parking disabled.

I am almost certain that you will find that the SB platform performs better with parking disabled. You can also test single threaded performance by writing a simple piece of C code that performs a MAC operation say...a few hundred million times without any SMP. Our system here in the lab gives almost a 2.5x increase in GFLOPS/s on the SB platform.

 

Well, the sandy bridge performs better with core parking disabled, whilst the clarksfield performs better with it on.
I find this weird considering they both quad core.

 

Yes but they are different architectures. Apart from the fact that they have four cores, there is nothing similar between them. SB was a HUGE upgrade over Nehalem.

Disabling core parking means you force all cores to work as hard as they can in effect, increasing heat and requiring more power. I'm almost certain that 840qm is throttling with core parking enabled.

Also, does the 840 perform better than the 2820 while both have core parking enabled?

 

Core parking simply means that the OS will not schedule a second thread on processor cores until all cores have already been awakened. The phenomena you are describing is simply because Clarksfield has very strict turboboost conditions and is extremely TDP constrained. When you allow corepark, you basically allow only one thread per core (or even park all cores if you tweak the Registry) so more resources are idling, this frees up TDP for the turboboost to work. However, with corepark off, there are fewer cores idling since the OS will schedule threads anywhere thus lower turboboost bins on individual cores.

Sandy bridge is not so strict with Multipliers (as opposed to TDP with clarksfield) due process node improvements and better turbo algorithms, it has a very high multiplier for all cores but with a TDP clamp so the cpu is better able to achieve burst tuboboost compared to Clarksfield.

As for single thread performance, don't forget Sandy Bridge is between 10-50% faster IPC vs clarksfield (depends on if you use unoptimized code or AVX instructions respectively), Sandy only needs to reach 2.9ghz to match the i7-840qm's single thread performance. And the higher multipliers coupled with better thermal headroom means that I am almost certain it will exceed the i7-840qm.

 

Yes this +1
Exactly what i am experiencing.
It feels like the turbo is better designed and peaks higher wiith 840 than 2820 especially in games eg crysis.
I will test 2820 with core parking on and see what peak freq it reaches in crysis 3. This will be interesting. Stay tuned!

 

How are you comparing this? You notice the stutter with parking enabled on the 2820 and you notice no or lesser stutter on the 840?

 

With crysis 3 and hwinfo64 sensors + cpu burn
Firstly the 840 suffers no stutter just it feels faster with core park enabled and this is proven with hwinfo plus cpu burn showing faster single clock peak freq.

The 2820 does stutter and to avoid core park must be disabled. Will test this with hwinfo and crysis3 soon and post if single cire peak freq is better/worst on core parking enabled.

 

May be you could test some other games as well? This could be specific to Crysis. Are you running the latest drivers? I read nv released drivers specifically for Crysis 3 (or may be not specifically but the game was the primary focus).

 

Ok just ran same tests on 2820QM.

With core parking enabled it hits 3392.4 Mhz on all 4 cores...max is 3400Mhz with cpu burn
Interestingly, with core parking disabled max is 3392.2 Mhz on 2 cores and 3392.4 Mhz on the other 2 cores again with cpu burn


Seems core parking is a unique process dependent on cpu architecture.
Going to game with crysis 3 and core parking enabled see if the stutter is there.

 

Yeah just played Crysis 3 and noticed very slight microstutter when walking forwards with core parking enabled. With it off it's much more smoother(hard to pick) and less obvious with Sandy bridge.

Edit: also looked at the clarksfield interestingly, when core parking is enabled the 2,3 and 4 core do not park as much as the sandy bridge. With sandy bridge you see the 1st core higher freq than the the other 3 cores at idle. With clarksfield they are more or less behave equally with core parking enabled at idle with peaks on the 3rd and 4th core whereas sandy bridge usually has the 1st and 2nd core peaking. Interesting to say the least and seems in some way they are reciprocal of each other. Moral of story sandy bridge disable core parking , clarksfield enable core parking. Enjoy

 

I have to add, you ARE using a different GPU with the Clevo than with the Alienware. That does make a HUGE difference and you can't just blame the CPU for being the problem here.