6MB L3 vs 8MB L3

[Lepton87]

Do you know of any tests that directly compare otherwise identical CPUs clocked at the exact same frequency with the only difference being 6MB L3 as opposed to 8MB L3? Ideally I'd like to see how it affects each of the Intel's recent architectures i.e SandyBridge, IvyBridge and Haswell. I wonder if there are differences in how each architecture benefits from the additional cache. Intel seems to put a very hefty premium on that additional cache. For example Core i7-3740QM is $378 and Core i7-3820QM is $568.00, I know these CPUs weren't launched at the same time so they aren't directly comparable pricing-wise, the CPU that's directly comparable is Core i7-3840QM which offers a mere 100MHz frequency bump and that additional cache all for 50% price hike. What adds more performance, the additional cache or those additional clock cycles? While 8MB L3 IB model also had 50MHz more in GPU turbo in HW that advantage, however small, goes away. Core i7-4810MQ 2.8GHz 6MB L3 $378,
i7-4900MQ 2.8GHz 8MB L3 $568 I would like to see those CPUs compared directly. I know it's mostly an academic comparison because the real upgrade is Core i7-4910MQ which is just a mere 100MHz clock bump over 4900MQ. Why does Intel price their fully functional CPUs so steeply? Is the performance increase from this additional cache more substantial than the performance increase from a mere 100MHz bump? I had 2500K so I could do some tests and then record the results and then compare them to my current CPU with HT disabled, all at the same frequency of course, but I didn't think of that at the time. Now I'm interested. I remember some benchmarks with a CPU that had L3 turned off entirely and then with L3 on all the way up to 8MB with 2MB increments. Does anyone remember that test? How do I turn off a portion of my CPU's L3?

 

[Chicken76]

Also interested in this.
I don't think you can just turn off a portion of the L3.

 

[Deders]

You could compare an i7 without hyperthreading to an i5 clocked to the same frequency.

 

[Keysplayr]

This might be obvious, but the extra 2MB of cache will only show advantages for programs or instructions that can make use of a greater than 6MB size cache.
Which may be zero programs. Way back when, when Celeron first hit the scene, I know they had zero cache or 128k. There was a huge difference in performance. Same with 128k vs 256k or 512k on Pentium II. but the difference was smaller.
I guess the answer is the same as it's always been. Depends on the usage.

 

[Jovec]

AMD example of 0MB L3 vs 4MB L3. It's likely that there are some benches out there with 6MB vs 8MB Intel i5 vs i7. Having any L3 at all is a good thing, but any difference between 6MB L3 and 8MB L3 is going to be exceeding small and only offer a benefit under very specific circumstances.


I think the answer to the OP is that if the CPUs both have L3, make a decision based on clock speed (or IGP performance or power consumption if relevant).

 

[Rakehellion]

The difference will be very tiny. Maybe about 1%-2%.

 

[cytg111]

really depends on usage, how many apps you running, what aps etc..

 

[Matt1970]

The difference will be very tiny. Maybe about 1%-2%.

I agree.

 

[Lepton87]

AMD example of 0MB L3 vs 4MB L3. It's likely that there are some benches out there with 6MB vs 8MB Intel i5 vs i7. Having any L3 at all is a good thing, but any difference between 6MB L3 and 8MB L3 is going to be exceeding small and only offer a benefit under very specific circumstances.


I think the answer to the OP is that if the CPUs both have L3, make a decision based on clock speed (or IGP performance or power consumption if relevant).

This was a purely academic question not about which CPU to buy. I guess I can envision a use case where the difference can be quite high like 10% or more, but for such a thing to happen a very important part of the program would have to be larger then 6MB but smaller than 8MB like 7.5MB. I imagine such use cases are exceedingly small, in general I don't expect the performance difference to break 2% but I wanted to see the exact results made with a consistent testing methodology as measuring such small differences in performance require a very small measuring error.

 

[dmens]

Also interested in this.
I don't think you can just turn off a portion of the L3.

Yes you can.

 

[Cerb]

This was a purely academic question not about which CPU to buy. I guess I can envision a use case where the difference can be quite high like 10% or more, but for such a thing to happen a very important part of the program would have to be larger then 6MB but smaller than 8MB like 7.5MB. I imagine such use cases are exceedingly small, in general I don't expect the performance difference to break 2% but I wanted to see the exact results made with a consistent testing methodology as measuring such small differences in performance require a very small measuring error.

Even over 8MB would do, so long as it doesn't get quite a bit large than cache. Outside of big server loads, and number crunching loads that need more CPUs cores anyway, though, that's likely somewhat rare. The general trend has been has been much more increasing the total number of smaller work units, than it has been greatly increasing each work unit's size. Where each work unit's size may be larger, like with photo or video editing, it's so much larger that even the L4 of CPUs equipped with Iris Pro gets dwarfed by the data. In those cases, the programs tend to be optimized to chop the work up so that most of what's done at a time fits at least into typical L2 caches, so big caches don't help much.

My guess is that having it all enabled on the K i5s is just to have one more way to differentiate them, since no benchmarks I've yet seen, nor trustworthy anecdotal evidence, either (multitasking improvements are hard to extract from benchmarks, sometimes), points to Core i5s with 6MB being any slower than 8MB ones at the same speed.

 

[cytg111]

Yes you can.

How ?

 

[dmens]

How ?

Whoops, didn't read the whole thread. Was thinking of another method but one that is not accessible to end users... sorry.

Anyways at sort when a tester determines where the erroneous lines are in the cache (if any) the part can be programmed to not use it. It's a way to recover yield. Consider the size of a last-level cache means a significant portion of dead dies will be due to defects there. This method creates useful parts out of what would have been junkers.

Related to OP's question regarding the steep price hike of functioning cache size is because die size = $$$. Same reason the disabling logic is there.

 

[cytg111]

Allrighty.
Now when you say "end users", does that mean that the ability has been permanently disabled in the final product or is there a theoretical way of playing with that stuff, even for us mere mortals ? Cause that would open op for some awsome benchmarking ...

 

[StrangerGuy]

The difference will be very tiny. Maybe about 1%-2%.

I remember one the quirks with Sandy Bridge L3 with synthetic cache benchmarks is that 6MB performs virtually the same as 8MB because the ring bus bandwidth became higher when the L3 becomes smaller.

 

[BrightCandle]

About the only way I can think of testing cache differences is to take two otherwise very similar CPUs and test them at the same clock speed. There is no way I know of to even turn the cache off on a modern CPU let alone just part of it, its simply not an option any bios current allows. You can do things like compare the 3960X to the 3930k and get 15MB v 12MB and you can do a 4770 with HT turned off and compare to a 4670. But AFAIK there isn't a way to do it with just one CPU.

 

[Cerb]

That's basically been done accidentally, just by having reviews of same-speed K and non-K CPUs w/ 6 and 8 MB L3.

 

[SPBHM]

the difference is small, I tried to google but didn't find with much, but, if you look on things like geekbench browser you can compare Sandy bridge e3 1220 and i5 2400, the exact only difference for performance is the l3 (6 and 8), clock and turbo is the same... with haswell e1220 there is no i5 with the same clock/turbo...

also if anyone have i5 4770K and 4670K it's easy to compare...

anyway, madshirmps compared the l3 but with only 2 cores working, 8 vs 6 vs 3.
http://www.madshrimps.be/articles/a...sary-Edition-G3258-CPU-Review/5#axzz39X1wpChU

 

[Magic Carpet]

In WinRAR, 4770K was about 8% faster at the same clocks, with HT turned off, last time I benched it against a 4570. But that's WinRAR and it loves cache, so yeah, depends on usage

 

[Grooveriding]

It depends on the application. Go back and look at Anand's SB-E review and look at the World of Warcraft results.

 

[cytg111]

Disabling it (L3) bit by bit would make an awsome log(n) graph per application. Imagine you could do the same with lower(higher?) level caches too. Awesomeness.