is there a difference in the l2 cache latency between allendales and conroes?

[grimREEFER]

the performance is just too close considering one has double the cache!
the only explanation i can fathom is lower latency L2 cache on the allendales
or i could be totally wrong....does anyone know?

[hicookie]

i test a abit with pi 1m, super pi 1m reflect cache , cache doesn't affect 32m



3.6G 2M VS 3.2G 4M

i think per clock/performance , L2 4m 20% faster than 2m

[hicookie]

super pi 32m and cpumark99 doesn't reflect cache issue

E6400@3.2G (8*400) 32M & cpumark99


X6800@3.2G(8*400) 32M & cpumark99

[Ace-a-Rue]

i read a review which showed an underclock 6800 at allendale speeds and then running benches between that and an actual Allendale cpu..there were only, the most, a 5-6% difference in bench performance and some were only 1-2%

[hicookie]

i read a review which showed an underclock 6800 at allendale speeds and then running benches between that and an actual Allendale cpu..there were only, the most, a 5-6% difference in bench performance and some were only 1-2%

really, can u show the link?

[Plaicd]

http://techreport.com/reviews/2006q3...f/index.x?pg=1

[grimREEFER]

yea, the conclusion they got was that there was barely any difference
thats what started me thinking
cause the situation now seems to parrallel when intel released the 2mb prescott and performed almost the same as the 1mb prescott

[Ace-a-Rue]

really, can u show the link?

unfortunately, i did not bookmark the review....sorry!

but, it is clear to me that L2 cache is important but the jump from 2 Mb to 4 Mb is probably not as significant as it looks on paper or feels in our mind.

[aashburns]

Have any websites looked at how the cache plays a role as the frequency increases? I'd like to see an X6800 with a 7x multiplier in comparison to a 6300 from stock to 2.8Ghz.

Also, does anyone know the associativity of the cache on the 2mb vs. 4mb? Is it only 4-way on 2mb or is it 8-way for both?

[grimREEFER]

look at the link plaicd posted lol
yea, the difference isnt as big as i think it should be, which leads me to believe the l2 cache on the allendale has a slightly lower latency than that of the conroe...it would be interesting to find out if this was true

[aashburns]

look at the link plaicd posted lol
yea, the difference isnt as big as i think it should be, which leads me to believe the l2 cache on the allendale has a slightly lower latency than that of the conroe...it would be interesting to find out if this was true

I looked at the techreport article earlier, but all they have is the overclocking page which does not give you any idea of how the cache size matters at higher clocks. What i'm saying is take an X6800, set the multiplier at 7. Then increase the fsb in increments along with an E6300 all the way upto a high clock such as 2.8GHz. Then compare how the cache influences the performance in various tasks. I believe multithreaded performance for encoding would be most interesting. So a line graph with performance deltas between the two processors. I'll look into other reasons for possible latency improvements on the 2mb. I believe that these are the same physical cores as the conroes with part of the cache disabled. The E4000 series will have a smaller reduced core. Even then I dont know if thats true because economically its cheaper to produce unused silicon and disable it than to make a separate mask and production process.

[Evilsizer]

Have any websites looked at how the cache plays a role as the frequency increases? I'd like to see an X6800 with a 7x multiplier in comparison to a 6300 from stock to 2.8Ghz.

Also, does anyone know the associativity of the cache on the 2mb vs. 4mb? Is it only 4-way on 2mb or is it 8-way for both?

allendale L1 and L2 is 8-way.

[Phosphate]

I did a thread on this. Apparently for 3D graphics Intel did not short change us L2 cache on the allendales (unlike my celeron where its L2 deficient)

As far as other applications go Anandtech reports anywhere from 3-10% improvement.

However I really want to know how 4 mb L2 cache scales in the high 3.5+ GHz clock speeds? Does it make more difference or less under those conditions?

[Evilsizer]

I did a thread on this. Apparently for 3D graphics Intel did not short change us L2 cache on the allendales (unlike my celeron where its L2 deficient)

As far as other applications go Anandtech reports anywhere from 3-10% improvement.

However I really want to know how 4 mb L2 cache scales in the high 3.5+ GHz clock speeds? Does it make more difference or less under those conditions?

if the clocks speeds are equal on both 2mb and 4mb the difference is still going to be the same as quoted above. if your running something like superpi for caculations all day long get the 4mb by all means.

[aashburns]

allendale L1 and L2 is 8-way.

Since Conroe is 16-way and if Allendale is 8-way, then the associativity of the cache is lower. Hence, there will be a slight performance hit from an increase in the rate of cache misses. I can't think of any reason why the allendale cache would have less latency is this case since it is the same die as conroe.

if the clocks speeds are equal on both 2mb and 4mb the difference is still going to be the same as quoted above. if your running something like superpi for caculations all day long get the 4mb by all means.

I suppose this depends on what you are doing. Remember that Memory Bandwidth does not increase proportionaly with the FSB (e.g. C2D can double its speed with careful overclocking, but memory cannot). Applications that frequently access more than 2MB data will definitely benefit from the larger cache. Not to mention at 3.5GHz the number of CPU cyles the C2D waits for memory (latency) is almost twice as much as a stock E6300. So the effectiveness of prefetchers is reduced as clock speeds increase. My guess is that Intel will include an on-die memory controller once the clock speeds reach a level at where prefetchers cannot hide the horrible latency of memory.

[xlink]

you have to realize that fullspeed cache is a lot easier to do without upping the latency on a lower clocked system than it is on a higherclocked netburst system. Atleast that is my guesse.

[Ace-a-Rue]

i read a review which showed an underclock 6800 at allendale speeds and then running benches between that and an actual Allendale cpu..there were only, the most, a 5-6% difference in bench performance and some were only 1-2%

really, can u show the link?

i found the answer and i posted it here:

http://www.xtremesystems.org/forums/...03#post1645603