Here's something that caught my eye from Hans de Vries over on Aces.

Johan wrote:

Interesting, seems more likely than HPEC's Square root theory to explain this :-)
There are two benchmarks which are specially Intel sponsored for the
occasion. DIVX and Valve's Half Life "Lost Coast" demo. These are the
two missing in HKEPC's 6400+/Conroe compare but included in the
Conroe/Penryn article:

http://www.hkepc.com/bbs/hwdb.php?ti...400&rid=842214
http://www.hkepc.com/bbs/hwdb.php?ti...iew&rid=837360

The Divx results come from the sum_of_absolute_differences instruction
for motion detection. This function should have been in the first MMX
instruction set in 1995 really. Now, in 2007, motion detection belongs in
the GPU in my opinion.....

As far as Half Life-2 is concerned: From all possibilities I assume that a
Last Coast set-size fit in the larger Penryn cache is mostly responsible
for the unusual improvement. I don't buy the shuffle arguments at all.

From extreme tech:
http://www.extremetech.com/article2/...2115086,00.asp

Half-Life 2: Lost Coast is a tech demo from Valve Software that's
actually a playable mini-game. A timedemo was recorded by Intel—
it's an actual gameplay recording rather than the canned graphics
performance test.

So, this allows the benchmarker, Intel in this case, to select a particular
recording in which the set-size/cache-fit gives a more than average
improvement going from Conroe to Penryn because it doesn't fit in 4MB
but does so in 6MB. It's just a question of playing the Last Coast demo
until you find the right scenes to record. Simple enough.


Regards, Hans
http://aceshardware.freeforums.org/v...r=asc&start=30


and

DavidC1 wrote:
Quote:
As far as Half Life-2 is concerned: From all possibilities I assume that a
Last Coast set-size fit in the larger Penryn cache is mostly responsible
for the unusual improvement. I don't buy the shuffle arguments at all.

Half Life 2: Lost Coast benchmark test system provided by Intel for the reviewers got 18.9% performance increase with 9.1% higher clock speed in favor of Penryn compared to Conroe. Despite a possible optimization that Intel MIGHT have done.

Half Life 2 benchmark tested by HKEPC showed that at equal configurations(same clock, same FSB), Penryn based core was 31% faster.

Your theory might have made more sense if HKEPC benchmark gained less, but it came out the opposite way.

You've got the wrong numbers (From the 3Dmarc06 test on the same page?)

The right numbers only confirm the hypothesis:

1.000 ==> 1.311 _____ HKEPC: (2,33 GHz/1.333 GHz Conroe DC==> 2,33 GHz/1.333 GHz Penryn DC)
1.000 ==> 1.359 _____ Intel:__ (2.93GHz/1.066GHz Conroe QC ==> 3.33 GHz/1.333 GHz Penryn DC)
1.000 ==> 1.391 _____ Intel:__ (2.93GHz/1.066GHz Conroe QC ==> 3.33 GHz/1.333 GHz Penryn QC)

http://www.hkepc.com/bbs/hwdb.php?ti...iew&rid=837360
http://www.extremetech.com/article2/...2115086,00.asp

Comparing the first with the last we see only a 6% improvement (1.391/ 1.311)
while the clockspeed improvement in the last case is almost 14% for the
CPU and 25% for the FSB so it's clear that the large improvement did not
come from a higher bandwidth or a faster CPU (higher frequency or faster
instructions) But way more likely from the larger cache.

The only difference between the latter two cases is an extra Die in the
package, QC vs DC, Still the relative improvement goes up from 3.5%
to 6%. Why would that be? Well how about having a whole 6MB L2 cache
available for the application while the OS mainly uses the 6MB on the
other die?

So by specifically recording the favorable scenes from the Lost Coast
demo where you find a large improvement (due to a scene size which
then must lay somewhere between 4MB and 6MB) you can create a
benchmark which shows a more than average improvement.


Regards, Hans.
http://aceshardware.freeforums.org/v...r=asc&start=45



Makes alot of sense. Is this possible?