Divider/performance question

Use my 4000+ as an example:
1:1 12x200 = 2400 (mem at 200)
2:# 8x300 = 2400 (mem at 200)

Will there be any difference in performance at all? Not just the standard fluctuation of benchmarks, but any real difference?

I would think the 8x300 would be faster

You may get some extra performance from the 300 setting, but why not test it out and see

Then let us know

Janus67 said:

I would think the 8x300 would be faster

Click to expand...

Actually, it's not faster.
If you use 3x HT multiplier then it's actually slower compared to 12x200. (3x300 = 900MHz compared to 5x200 = 1000MHz) A64 based system gains nothing in name of performance if higher HTT is used.

Speed of the CPU is more to worry about here. The higher your CPU is clocked, the more the performance you gain, also the higher your CPU clock frequency is, the more the memory bandwidth you get out of your memory.

I'd stick on 12x200 with 1:1 memory setting and do some overclocking


Back to time when I had MSI K8N Neo2 Plat and A64 3500+ NewCastle, I tried out 10x240, 8x300 and 7x343. No performance gain out of it, so I thought that why should I stress my system more with higher HTT speeds if I can't get anymore performance out of it than I do with lower HTT speeds.

But, if you like higher numbers, then do so, but you see no extra performance gains

So, the name of the game is to still get HTT/FSB as high as stable, find the max your CPU can go, then find some combination of dividers/multipliers that will allow you to have both. Is that what I'm reading here?

Kil4Thril said:

So, the name of the game is to still get HTT/FSB as high as stable, find the max your CPU can go, then find some combination of dividers/multipliers that will allow you to have both. Is that what I'm reading here?

Click to expand...

Well, yes and no.
The max HTT is not name of the game anymore, not with A64 system, but with AXPs and P4 systems it is. Higher HTT doesn't give you advantage over lower HTT, but the clock frequency of your CPU does. Faster you can get your CPU clocked, the more performance you have in use in it's memory controller. Examples:
12x200HTT = 2400MHz
10x240HTT = 2400MHz
8x300HTT = 2400MHz

All the examples I just wrote, are all in same line, 300HTT is not any faster than 240HTT or 200HTT (in case that memory would run all the time at 200MHz).

And now, back to the basics:
First you should try and take the maximum clock frequency out of your CPU, after that try the memory how far it can go. After you have done both of these, find a matching combination for them. Example, if you have memory which can do max 270MHz and you have CPU which does something like ~2720MHz, just take 10x multiplier to use and crank the HTT upto 270MHz so you can get your CPU running at 2700MHz and your memory 270MHz.

There's also this thing called "HyperTransport Bus", but it's really something you should not be worried about cause it effects so slightly to performance.

Sorry if my text is somewhat confusing or anything, it's 5:26am at here so I think I'll go to bed now. If you have questions, just ask, I'd be more than glad to help you out

Actually, Zebbo, you said exactly what I thought I was trying to say. Time to start pushing this Clawhammer- I just hope I can cool it in the case it's currently in (little modification going down).

Zebbo said:

Actually, it's not faster.
If you use 3x HT multiplier then it's actually slower compared to 12x200. (3x300 = 900MHz compared to 5x200 = 1000MHz) A64 based system gains nothing in name of performance if higher HTT is used.

Speed of the CPU is more to worry about here. The higher your CPU is clocked, the more the performance you gain, also the higher your CPU clock frequency is, the more the memory bandwidth you get out of your memory.

I'd stick on 12x200 with 1:1 memory setting and do some overclocking


Back to time when I had MSI K8N Neo2 Plat and A64 3500+ NewCastle, I tried out 10x240, 8x300 and 7x343. No performance gain out of it, so I thought that why should I stress my system more with higher HTT speeds if I can't get anymore performance out of it than I do with lower HTT speeds.

But, if you like higher numbers, then do so, but you see no extra performance gains

Click to expand...

I would respectfully sayYou got it wrong.
4x300=1200mhz. notice 1200 mhz. is the final speed of cpu now lets see another way 6x200=1200mhz final cpu speed same as above, but the big difference is the bandwidth 4x300 is higher bandwidth than 6x200. Another word w/ the first option you have reduced the cycle but increased the burst(electromagnetic charge). 3-220 plugs has more power to it than 6-110, interms of electricity. same is w/ cpu and bus burst.

ochungry said:

I would respectfully sayYou got it wrong.
4x300=1200mhz. notice 1200 mhz. is the final speed of cpu now lets see another way 6x200=1200mhz final cpu speed same as above, but the big difference is the bandwidth 4x300 is higher bandwidth than 6x200. Another word w/ the first option you have reduced the cycle but increased the burst(electromagnetic charge). 3-220 plugs has more power to it than 6-110, interms of electricity. same is w/ cpu and bus burst.

Click to expand...

Overclocking about the HyperTransport Bus has been in discuss much earlier and almost everyone has come to conclusion it's not worth to use lower CPU multiplier and higher HTT (FSB) in A64 platform. In A64, the ruler is the frequency of CPU core speed.

I did some benchmarking to gave some proof to what I was talking about.

First screenshot here
CPU: 1808MHz (226x8)
Memory: 200,9MHz (2-3-2-6)
HTB: 1130MHz (5x226)

My SuperPi8M time was 8minutes 15.703seconds

Second screenshot here
CPU: 1806MHz (258x7)
Memory: 200,7MHz (2-3-2-6)
HTB: 1290MHz

And SuperPi8M time was 8minutes 15.578seconds.

Let's do the conclusion then.
With 160MHz higher HTB the gain was only 0.125seconds in SuperPi8M time. Not much of an improvement. With 200MHz higher HTB, the gain could be in SuperPi something like 0.13seconds, and probably not even that. If you want more results, google will help you out, just type "HTT Overclocking".