Cache - my 512k or 1M delima.

[dedla]

4200+ & 4600+ are E4's, 4400+ & 4800+ are E6's - Is there any difference besides cache sizes and speed increases? Why bother giving them different codes, unless it's the easier way of saying Manchester and Toledo?

Now for the big cache problem; I keep reading the 1M L2 doesn't make much of a difference in gaming, or general standard performance comparared to 512k of L2. WHY doesn't it make a difference? If it makes no difference in gaming or general apps, where DOES 1M of L2 matter? Why have 1M available at all? Is there any chance or place in gaming where it might make a difference now or in the future? Would it even theoretically help with 64 bit OSs?

Articles explaining or showing these answers are welcome, for I couldn't find any after searching the net for 4-5 hours last night.

 

[mpjesse]

The reason there isn't much of a difference is because there's a hit in latency when you increase the size of the L2 cache. (this was discussed in the Pentium D 955 review on this site)

Currently, both AMD and Intel processors can address and speak to smaller L2 cache sizes faster than large L2 caches.

So it's a catch 22. The processor can do more when it has a bigger cache (1MB per core), but it takes longer to do more. Vice versa w/ 512Kb caches.

Make sense?

-mpjesse

 

[ShortyNumber45]

Theoritically, IMO CPU's w/ lesser cache size performs better in games while CPUs w/ more cache(i.e. 1M or 2M) performs better on none gaming applications.

 

[RichPLS]

If there is very little difference, why does AMD make an X2-4200 2x512k cache and an X2-4400 2x1MB cache yet both are at same clock speed of 2200Mhz, sounds as if the extra 2x 512k or cache can make around 9% boost, enough to brand it so.
What about a chip with 0 L2 cache, remember ye ol' celeron of yesterday?

 

[dedla]

Thank you all for the quick responce, yes it makes a bit more sense now. Was worth waiting most of a day to sign up for the forums instead of making my eyes bleed trying to find more answers on sites that didn't exist.

I shall be checking out the Pentium D discussion now. I saw hints of longer access times but only heard them mentioned to Pentiums, was hoping AMD didn't ahve those problems.

I was under the impressions Celeron's were dispised becasue of their lack of L2, anyone who knew anything knew that a mere 128k L2 made Celeron's wimper and beg like a whupped pup.

 

[RichPLS]

Celerons have cach now, that was then.

 

[ltcommander_data]

I agree with wusy that more cache should be faster in games. I believe the explanation has something to due with the ability to fit large or more textures into the cache allowing faster processing making size and bandwidth more critical than latency. Everyday applications, especially Office, benefits from lower latency rather than bandwidth on the other hand because the files are generally smaller and don't require such a large cache.

It also isn't true that larger caches are always higher latency. On the Intel side, the Celeron D and 5xx models all share a common latency of 23 cycles despite the fact that the Celeron D has 256k of L2 to the 5xx's 1MB. The 6xx (2MB) and 8xx (1x1MB) likewise share a common latency slightly higher common latency of 27 cycles. As well, despite what mpjesse has said about the increased latency of the 2x2MB cache of the 9xx over the 8xx it isn't true. The L2 cache of the 9xx has been doubled while maintaining the same 27 cycle latency. The 8xx always had the added latency of the 6xx series despite it having less cache per core. The new 65nm Celeron D's coming out will have double the cache to 512k and I presume the latency will be increased to 27 cycles to correspond to the 6x1 series they are based on.

http://www.anandtech.com/cpuchipsets/showdoc.aspx?i=2658&p=4

The most telling example of when larger caches can be obtained without added latency is the Pentium M. Dothan doubled the L2 cache of Banias to 2MB while still maintaining the low 10 cycle latency. As well, the reason why AMD can quote an extra 200MHz in the PR rating for doubling the cache is because their increase is likewise obtained without increased latency. I believe both 512k and 1MB L2 cache A64s have a latency of 17 cycles.

Generally though, the extra cache doesn't make a significant difference in performance. Only applications that can readily fill up such a large cache show a performance difference. This is especially true for AMD's architecture which, due to the use of an exclusive cache, doesn't rely as heavily on the L2 cache and so don't see as large an increase in performance when it is increased. A large part of it is also that the point of a L2 cache is to increase hit rates to reduce the likelihood of going to RAM, but once the hit rate is say 99.99% (just making up the number to illustrate) an extra 0.009% won't make as much of a difference.

I believe that endyen usually said that the extra cache is useful as a heat sink so if you are overclocking, the extra cache is a good idea. In any case, MHz provides a more reliable performance boost.

Oh, and about 64-bit. I'm going out on a limb here since I'm not completely sure, but I believe that extra cache will be more helpful in that case since the instructions are longer (twice or less?) and so would take up more cache space.

 

[mpjesse]

I stand corrected. Thanks for pointing this out Data.

-mpjesse

 

[dedla]

Your article about anandtech's test is just making me laugh at Intel more.

I probably should've pointed out earlier I'm looking for a long term investment with this discission, with absolutely no OCing in mind. I'm looking into this on the basis of future security with 64-bit processing and/or OSs. I lasted 5-6 years fine on my P3-500Mhz, which I'm still debating turning it into a legacy machine for Win98-stable games, or whether or not I should just try to virtual machine them on a newer system.

I'm looking into a processor and set up that'll last a comparible ammount of time. The 4800+ from your article's looking mighty sexy now, but what's 200Mhz...almost feel I should be starting another board for this, bit I KNOW this topic's covered in other places.

Kinda off-topic, everyone's loving Opterons for the overclockability, but I noticed thy don't have near as many CPU command features, exp. SSE3 (not sure if it's what they're called). Anyways, how useful is overclocking the crap out of an Opteron when the X2s have more diverse and updated command...thingies? Couldn't they be doing the same operations in less cycles?

 

[ltcommander_data]

I believe the 1xx Opterons are just higher binned A64s and X2s for lower voltages and better stability. As such Opterons would have all the same features and instruction sets, if not more, then the regular consumer products. They may not advertise it, (in fact they prefer not to advertise Opterons to consumers who've created a supply crunch for corporate customers), but SSE3 is there.

If you are choosing between a regular X2 or A64 and an Opteron you might as well get an Opteron. Even if you don't overclock if won't hurt you any, and you'd have the assurance of extra reliability. It's also nice to tell people you have an Opteron too. You should note that Opteron prices are going up soon though to discourage regular consumers from buying them.

 

[Admiral_Cecil]

Hello,

Opteron's for the 939 have all the features X2's have. Same with the single core and Opteron equiv.

Opteron's are like FX, they are the best of the single core and best of the dual core. Originally for the server, now even the average person can buy them if they know what they are looking for.

Now that's not to say 3700+ and 4400+ are bad performers at all. But Opteron's give you FX-like quality for less than an FX.

 

[dedla]

Well, you and Data settled it for me, I'm going after and Opteron 175. The 180s are just insane for just a couple hundred more MHz.

Thanks again for all the help guys.

 

[mito]

I ordered my Opteron 175 last week, should be arriving anytime soon...

 

[pat]

I shall be checking out the Pentium D discussion now. I saw hints of longer access times but only heard them mentioned to Pentiums, was hoping AMD didn't ahve those problems.

Problems.. that aint problem.. that normal.. would it take longer for you to look for something in a small full box than a quite large box. and a small box might nor even have it soo..

Too low on cache dont help. just enough is perfect. too much is not better.

Just have to know what too little is, what just enough is and what too much is. But there, it depend on the applications.

Game, with their impredictible movement make cache less usefull, because this is the player that mostly decide what's coming next. In office apps, well, usually sit there waiting for user input or playing solitaire, so cache would be not that much stressed.. but rendering and compressing, with their repetitive instructions will benefit from more cache.. but not from too much thou..

 

[noko]

I have an Opteron 165 running at 2.43ghz and couldn't be happier. Limited by motherboard and not by cpu. The Asrock 939 Dual SataII newer bios's has a 274 cap on frequency. The older bios's or OCWbeta which doesn't have this limitation don't run on my machine with 1T timings. Anyways Opteron's are great from my experience so far.