512KB versus 1MB L2 cache

[Wet Fart]

I realise this discussion has been had many times over on this forum but I'm now going a bit cross-eyed doing thread searches to find threads containing information on this question.

I belong to a forum in Australia called Whirlpool ( http://forums.whirlpool.net.au/forum-threads.cfm?f=7 ) and we see a great number of new userslooking for guidance, and it would be a terrific help to me to understand exactly what types of applications benefit most from the larger L2 cache.

This will enable our users to be able to make a wiser decision, when considering the cost versus performance factors that the the new San Diego cores and X2 processors present.

Thanks for any links and discussion you would like to share

[cirthix]

i'd say less than that, maybe 0-5% depending on the application (normal apps here). maybe up to 15% in synthetic stuff.

[Primoz78]

Have tried the Venice 3000+ @ 2,7ghz and the SD 3700+ at same frequences...in Super_pi, sandra,...and have the same score...+/- max. 2-3%...in most cases the difference was a + on the venice side!

Regards,
Primoz

[Wet Fart]

Thanks guys.

I'm wondering what types of applications would be L2 cache intensive, some way to be more specific, to bench any differences.

I mean there must be some greater advantage to having twice the cache, that will see more than 2-5% increase.

A good start, what do you think would be a good method of benching L2 cache, specifically.

This is very interesting

[El Snorro]

3d benchies (especially 2k1se) just love 1mb l2 cache.

So if youre a 3d bencher 1mb is the way to go

[SIOUX]

Have tried the Venice 3000+ @ 2,7ghz and the SD 3700+ at same frequences...in Super_pi, sandra,...and have the same score...+/- max. 2-3%...in most cases the difference was a + on the venice side!

Regards,
Primoz

Thats hard to understand..... My 3700+ was atleast 2 sec faster, than my 3000+ venice at same clocks in Spi1M....

[Wet Fart]

Hey, what's happened to your avatar SIOUX?

It's stopped looking all trippy......or......did I finally come down?

[Primoz78]

Thats hard to understand..... My 3700+ was atleast 2 sec faster, than my 3000+ venice at same clocks in Spi1M....

I know...but that was the facts...I was very disapointed with it so have gone for an X2...

Regards,
Primoz

[Ugly n Grey]

250Mhz? I might swallow 100Mhz everything being equal... but 250Mhz?

[twiggy]

250Mhz? I might swallow 100Mhz everything being equal... but 250Mhz?

I think its gotta be ~200mhz as the 3400+ NC is 2.4ghz 512kb L2 and the 3400+ CH is 2.2ghz 1mb L2. Not sure how they compete with each other but I could run some Pi tests because my bro and me have the almost same rigs excepts he has a 3400CH and I have a 3400NC.

[GRAFiZ]

I've recently had a chance to play with this very subject and have come to simple conclusions.

Basically, I used an X2 4400+ /1mb cache overclocked to 2.6ghz. And in the same system a Venice 3000+ /512k cache overclocked to 2.6ghz.

Now then, in both cases I loaded up NeroVision Express and in the case of the X2 set affinity to CPU 0 only.

I then took three video files, roughly 2.2 gigs on size and set on to encoding for DVD.

On the X2 (affinity at CPU 0) the time to encode was right at 72 minutes.

On the Venice at the same speed it took right at 76 minutes.

So, 4 minutes difference.

I'm terrible with math, but what are we talking 4% to 5% difference.

Hardly anything to write home to your family about. But, it would account for perhaps 100mhz in speed?

Hard to say directly really as my test might be a touch tainted by a dual core CPU, but, like I said I forced CPU 0 and watching the task manager the whole time never saw a bit of useage on the second core.

No big deal really, but hell, 5% is 5%.

EDIT - Let me add, in both cases everything in the bios was at default setting except for the FSB which in both cases was set to the same speed. 9x288.

If I were to play with HTT and memory timings that difference might go one way or the other.

[J-Mag]

In new generation games at high resolutions the cache doesn't even help at all.

[GRAFiZ]

In new generation games at high resolutions the cache doesn't even help at all.

Yeah, I found the same when running games I could not find more then a frame or two difference between the two.

However, if you do video or graphics work, the difference is there, but, worth the price? I dont know, up to the buyer I suppose.

Now, when we can get 8mb cache's we'll really have something!

[J-Mag]

Yeah, I found the same when running games I could not find more then a frame or two difference between the two.

However, if you do video or graphics work, the difference is there, but, worth the price? I dont know, up to the buyer I suppose.

Now, when we can get 8mb cache's we'll really have something!

Yeah I agree. I can imagine that one reason why we don't see too much improvement with PC games is because most developers were writing code when the only friggin AMD proc with 1mb was the FX series...

[SIOUX]

Hey, what's happened to your avatar SIOUX?

It's stopped looking all trippy......or......did I finally come down?

LOL....for me it looks normal here??? anybody else dont see the avatar???


hehe....maybe you did not come down from the shrooms yet

[J-Mag]

LOL....for me it looks normal here??? anybody else dont see the avatar???

I see it but your animation is gone...

[Ugly n Grey]

No big deal really, but hell, 5% is 5%.

.

Right, but affinity was only set for the application while the other CPU was free to share background resources for Windows.... so it's not an equal test, you need to test two procs with everything the same except the cache...

[XTCgr]

you can check some tests i did http://www.thelab.gr/showthread.php?...n+diego+venice

it is in greek but the language of numbers is universal

[bigjohns97]

http://www.anandtech.com/cpuchipsets...spx?i=2330&p=2

this will kinda give you an idea on a cpu limited real world game.

[Ugly n Grey]

I am still kind thinking 100MHz is all it would take to catch the bigger cached chip....none of these numbers are that big that that wouldn't be enough

[i found nemo]

128k compared to 1024k is about 4-6 sec in pi....if anyone cares

[GMX]

The performance variaition between 512kb, and 1mb at higher Frequencies should be higher rate of change than at lower speeds.
For example at low speeds 512kb vs 1mb could be 200mhz. At higher speeds 512kb could need more than 300mhz to catch up.

[Ugly n Grey]

The performance variaition between 512kb, and 1mb at higher Frequencies should be higher rate of change than at lower speeds.
For example at low speeds 512kb vs 1mb could be 200mhz. At higher speeds 512kb could need more than 300mhz to catch up.

benchies?

[Wet Fart]

Thanks for the input guys

It's a start.

[Nano2k]

Check this link out, not much real world difference between 128k and 1024k cache imo.

http://www.xbitlabs.com/articles/cpu...on-2600_5.html