mersenneforum.org  

Go Back   mersenneforum.org > Fun Stuff > Lounge

Reply
 
Thread Tools
Old 2002-08-23, 17:16   #34
xtreme2k
 
xtreme2k's Avatar
 
Aug 2002

AE16 Posts
Default

Angular:

You are correct on the reduced latency when you run 333MHz memory speed on a 266MHz FSB. This is bloody obvious :) Anyway, however, this still isnt the optimal way to run memory at 333 since now the FSB and RAM is now in Asynchorus mode therefore 'some' latency is actually added to 'make up' for the difference in speed. However, after adding such asynchous latency, it is still slightly lower compared to 266FSB:266RAM. Plus you have 266MHz FSB bottleneck, that is why KT333 never actually blew away the performance of KT266A chipsets.

For BEST performance, it is best if you unlock the CPU and run BOTH the FSB and RAM at 333. This way you get around 5-10% of extra performance AT THE SAME CPU MHz because now the latency is greatly reduced, no async latency needed to be added, as well as the FSB itself having alot more bandwidth.
xtreme2k is offline   Reply With Quote
Old 2002-08-23, 18:30   #35
ebx
 
ebx's Avatar
 
Aug 2002

101 Posts
Default

Well, XP and P4 can both do prefetch. If the code is optimalized that way for the critical path, you will always have your data in cache when you need it. Prefetching gave me 30% boost when I implemented it for a work project. It may even be better for todays P4/XP since their frequency is a lot higher and memory access speed still in the same order. 2G CPU clocked at 0.5ns and 133FSB memory cycle is (I am guestimating) still at over 10ns. (Hey, maybe I can do that for prime95 and claim a slice of the EFF prize. :D ) You dont know how your bus is utilized until you put it under a logic analyzor. Besides, all the memory access are bursted meaning you get a whole cache line in even if you only need a byte from it. If your data access folllows the locality anticipated, by bursting your next data is in cache now. XP has a 64 byte L2 cacheline. That further reduce the latency effect of async buses.

I didnt read kt333 spec. Dont know if data is buffered. I admitted for the chips I worked with, they were all sync.

Angular, I will sent you a PM for the vendor. It is not a big vendor. Boxes from big vendors can never be performance king. They may offer low cost and good services for volume buyers. But never performance in our sense.
ebx is offline   Reply With Quote
Old 2002-09-06, 20:19   #36
sdbardwick
 
sdbardwick's Avatar
 
Aug 2002
North San Diego County

2B116 Posts
Default

Quote:
Originally Posted by ebx
Prefetching gave me 30% boost when I implemented it for a work project. It may even be better for todays P4/XP since their frequency is a lot higher and memory access speed still in the same order.[...](Hey, maybe I can do that for prime95 and claim a slice of the EFF prize. :D )
Actually, Prime95 has utilized prefetch since version 21.2 :D
From the whatsnew.txt file:
Quote:
2) The program now uses the prefetch instructions on the Celeron 2, Pentium 3, and all Athlon CPUs. This results in about a 20% performance boost on these machines.
-Scott-
sdbardwick is offline   Reply With Quote
Reply

Thread Tools


Similar Threads
Thread Thread Starter Forum Replies Last Post
Factoring database Syd FactorDB 1627 2021-06-22 16:14
Database for k-b-b's: 3.14159 Miscellaneous Math 325 2016-04-09 17:45
CPU Performance Database JustinGC Hardware 3 2012-06-22 03:11
NPLB Database IronBits No Prime Left Behind 177 2009-10-10 09:00
database.zip HiddenWarrior Data 1 2004-03-29 03:53

All times are UTC. The time now is 06:10.


Sat Oct 23 06:10:20 UTC 2021 up 92 days, 39 mins, 0 users, load averages: 1.51, 1.41, 1.33

Powered by vBulletin® Version 3.8.11
Copyright ©2000 - 2021, Jelsoft Enterprises Ltd.

This forum has received and complied with 0 (zero) government requests for information.

Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.2 or any later version published by the Free Software Foundation.
A copy of the license is included in the FAQ.