2MB L3 cache vs. 1MB L3 cache

opuscroakus · February 1, 2003 11:31PM

The new dual 1.42 has 2MB L3 cache per processor vs. the 1MB L3 cache per processor on the dual 1.25. How much of an extra boost will the 2MB L3 cache on each processor provide over the 1MB cache?

Will this extra L3 cache goodness really help or is this more of a marketing thing. <img src="graemlins/hmmm.gif" border="0" alt="[Hmmm]" /> <img src="graemlins/hmmm.gif" border="0" alt="[Hmmm]" />

coolmac · February 1, 2003 11:43PM

I remember in the past reading that the extra MB of L3 cache can be very important, perhaps almost as important as the processor upgrade It definitely is supposed to make a difference.

bunge · February 2, 2003 12:15AM

In all honesty I've heard that 2MB sometimes causes a reduction of speed. Try Ars, that might be where I saw it. If not there maybe macspeedzone or some similar site. It was over a year ago probably, but I remember it clearly since the tests came up with such an odd conclusion (just for the record, I think it was slower because when a refill is necessary it took a lot longer to fill the cache with data.)

roduk · February 2, 2003 8:41AM

I believe the size of a cache can reach a point where the overhead in determining whether an instruction or data is present in the cache, finding it isn't and fetching it from main memory into the cache is larger than if you had just fetched it from main memory in the first place. I'm not sure what the optimum cache size is though.

[ 02-02-2003: Message edited by: RodUK ]

rogue27 · February 2, 2003 10:36AM

It depends on the software you are using. There shouldn't be a large cache speed penalty if the software developer is smart and tells the processor to fetch certain data from main memory without looking in the cache first.

For example, if you were rendering video and the processor was instructed to fetch each new frame of video from main memory, then there wouldn't be a speed penalty. If the developer left that code out, the processor would, by default, search the cache first and then look in main memory if it needed to. This is obviously going to be slower than skipping the cache if you *know* the data won't be in the cache.

Code to ignore the cache like this would probably be left out in most cross platform software because most PCs have small or nonexistent L2 or L3 cache so most developers wouldn't add such instructions to their code. Thus, when they do cross-platform development, they don't add it on the mac version because it's not there on the PC version and they are often less knowledgeable or concerned with the mac version of their software.

Software written to run well on a mac can run fast. Mac software that is written like PC software will run slower. The two architectures have many differences and if developers don't take them into account then the hardware won't get used to it's fullest ability.

As to the 2MB or 1MB discussion... just wait for benchmarks comparing the new 1.25Ghz machine with the old 1.25Ghz machine and see how they stack up. However, I wouldn't necessarily trust anything on barefeats.com.

[ 02-02-2003: Message edited by: rogue27 ]

g-news · February 2, 2003 10:41AM

following your logic that more cache means slower performance, an old athlon would have to be faster than the new ones, and a POWER 4 with 128MB cache would be the slowest chip on earth.

Since that is clearly not the case, that theory is way oversimplified.

It would be interesting to test an old Dual 1250 G4 vs a new one...the only difference is the cache (2MB vs 1MB)

G-News

the cool gut · February 2, 2003 11:26AM

here's the article at speedzone:

<a href="http://www.macspeedzone.com/archive/5.0/upgrades/G4350-2MB-XLR8.html"; target="_blank">http://www.macspeedzone.com/archive/5.0/upgrades/G4350-2MB-XLR8.html</a>;

gardenofearthlydelights · February 2, 2003 3:12PM

This is a link to <a href="http://www.powerlogix.com/press/releases/2002/020826.html"; target="_blank">Powerlogix's white paper</a> on L3 caches. It was discussing the difference between DDR and SDR memory, but somewhere they state that Bigger is Better, as we all know.

stoo · February 2, 2003 4:34PM

Cache is (mostly) transparent to processes: hardware deals with cache hits/misses rather than the code. Code can still be optimised for caches of specific size, speed, etc...

smalm · February 3, 2003 7:47AM

[quote]Originally posted by opuscroakus:

Will this extra L3 cache goodness really help or is this more of a marketing thing. <hr></blockquote>

The bus ratio for the DP 1250 is 7.5 and each processor can only use the bus half the time (for read, not for write). Throw in the much higher latency for main memory vs L3 cache and you can see how much higher the penalty a cache miss in the L3 for the DP 1250 vs a cache miss in the L2 for the SP 350 is.

Apple clearly should build in 2 MB L3 per proc in the whole pro line.

February 3, 2003 11:35AM

[quote]Originally posted by the cool gut:

here's the article at speedzone:

<a href="http://www.macspeedzone.com/archive/5.0/upgrades/G4350-2MB-XLR8.html"; target="_blank">http://www.macspeedzone.com/archive/5.0/upgrades/G4350-2MB-XLR8.html</a><hr></blockquote>;

I don't know what this Ron Skoog

was smoking but everything he says is complete BS.

airsluf · February 4, 2003 4:32PM

serrano · February 4, 2003 10:49PM

***CONFIRMED***

Apple crippled the 1.42 with 2MB of L3 Cache to increase sales of the 1.25!

OH NOES!#@!!

Seriously though, that extra MB per processor is a good thing.

2MB L3 cache vs. 1MB L3 cache

Comments