Dá que pensar heim?
"Thanks to Sampsa Kurri from the Finnish site Muropaketti.com for posting about the Pentium 4 iTLB. Apparently, the Pentium 4 has always had 128 instruction TLB entries, and not just the new stepping as reported in our review. This information is confirmed on page 364 of the IA-32 Intel Architecture Software Developer's Manual, Volume 3:
Pentium 4 and Intel Xeon processors: 128 entries, 4-way set associative.
P6 family processors: 32 entries, 4-way set associative.
Pentium processor: 32 entries, 4-way set associative; fully set associative for Pentium processors with MMX technology.
According to Sampsa, there is a bug in older revisions of the Pentium 4 that mistakenly reports 64 entries instead of 128. This has been fixed in the newest stepping, causing WCPUID to report the correct value.
My friend told me that there is a bug in older Pentium 4 CPU steppings. EAX=0000_0002h register returns wrong value. (If Intel document is correct and there should be 128 entries).
If you look at this document and page 8:
http://www.intel.com/design/processor/future/manuals/Cpuid_supplement.pdf
This document tells what Descriptor values are for 64 entries, 128 entries and 256 entries. For some reason in older P4 stepping (B-0) EAX=0000_0002h returns value 50h but it should return value 51h.
So, the question is, if the TLB has not been changed, then what's responsible for the per-clock performance improvement of the new stepping? Upon contacting Intel, Sampsa was told there were no performance improvements in the new stepping, despite the fact that he has reported a 10 second improvement in SuperPi at the same clockrate. Indeed, our own benchmarks have indicated a 1 to 5 percent IPC increase in several benchmarks. We saw a 7 second improvement in ScienceMark 2.0's Moldyn benchmark and a 28 second improvement in Primordia, at the same clockrate.
We've contacted Intel for more information and we hope to have a response sometime today. As soon as we have it, we'll update both this post and our review with details."
http://www.aceshardware.com/
Pois é... pensava-se que Intel tinha mudado o steping do CPU para só alcançar mais velocidade, e que tinham aumentado o tamanho dos TLBs como aconteceu da passagem do TBird para o Palomino, e afinal era por causa do WCPuid que fazia o report errado dos P4 relativo aos TLBs.
Mas algo a Intel mudou para que o CPU desse mais clock a clock em relação aos "manos" mais velhos...
E esta hein?