Testing / comparing : Intel D975XBX2 / Asus P5B DX ***56K WARNING***

[gtj]

Wonder if Penryn will work in the BX2.

I would think "yes" with a BIOS update but I guess we won't know until Intel makes more info available.

[argylesocks]

A 1:1 FSB to Memory ratio sounds nice but the critical comparison is FSB throughput to real-world memory throughput. Ideally, you want the 2 to be close.

can you expand on this a bit, gtj?

the whole "1:1 ratio" was drilled into my head as the "best" when i started here... but a few brave souls said they thought a 2:3 or 4:5 gives better results...

care to shed some more light?

fyi: im running a 4:5 ratio now, and dont really notice a difference either way.

[gtj]

can you expand on this a bit, gtj?

the whole "1:1 ratio" was drilled into my head as the "best" when i started here... but a few brave souls said they thought a 2:3 or 4:5 gives better results...

care to shed some more light?

fyi: im running a 4:5 ratio now, and dont really notice a difference either way.

I may have been one of those who said 1:1 is the best but after doing the math and running benchmarks, I've changed my mind and may have to apologize for making statements without evidence to back them up.

1:1 sounds nice. The FSB and memory are running at the same clock speeds so that must be the most efficient, right? All other things being equal, maybe 1:1 is more efficient that 4:5 but the reality is somewhat different.

The fact is that any "efficiency" the MCH may see is vastly overwhelmed by another fact.... Data passing between the processor and memory is limited by the SLOWEST link. If your FSB is running at 8 GB/s and your memory bus is running at 6 GB/s, then data can only pass between the processor and memory at 6 GB/s. PERIOD. If moving your ratio from 1:1 to 4:5 brings the memory bus to 7 GB/s that's a 17% increase. Gee, so it may not be as efficient as 1:1. Who Cares? Honestly, I can't even find any documentation to support that there was any efficiency loss in the first place.

This is easy to test. Set your FSB to a value that let's you run 1:1 plus either 2/3 or 4/5 reliably. Run an Everest Cache and Mmeory benchmark at 1:1, then without changing anything else, set a 2/3 or 4/5 and run the benchmarks again. Throughput should go up and latency should go down. What a concept.

[ChaosMinionX]

I may have been one of those who said 1:1 is the best but after doing the math and running benchmarks, I've changed my mind and may have to apologize for making statements without evidence to back them up.

1:1 sounds nice. The FSB and memory are running at the same clock speeds so that must be the most efficient, right? All other things being equal, maybe 1:1 is more efficient that 4:5 but the reality is somewhat different.

The fact is that any "efficiency" the MCH may see is vastly overwhelmed by another fact.... Data passing between the processor and memory is limited by the SLOWEST link. If your FSB is running at 8 GB/s and your memory bus is running at 6 GB/s, then data can only pass between the processor and memory at 6 GB/s. PERIOD. If moving your ratio from 1:1 to 4:5 brings the memory bus to 7 GB/s that's a 17% increase. Gee, so it may not be as efficient as 1:1. Who Cares? Honestly, I can't even find any documentation to support that there was any efficiency loss in the first place.

This is easy to test. Set your FSB to a value that let's you run 1:1 plus either 2/3 or 4/5 reliably. Run an Everest Cache and Mmeory benchmark at 1:1, then without changing anything else, set a 2/3 or 4/5 and run the benchmarks again. Throughput should go up and latency should go down. What a concept.

Setting say FSB to 400... Then running 266ref and mem at 533 gives you DDR2 800 1:1 but if you set 667 it gives you ddr2 1000 4:5 and if you keep the same timings and FSB of course its going to be faster and have lower latency you changed the memory frequency to 500 instead of 400. Or am I reading what you said wrong?

[gtj]

Setting say FSB to 400... Then running 266ref and mem at 533 gives you DDR2 800 1:1 but if you set 667 it gives you ddr2 1000 4:5 and if you keep the same timings and FSB of course its going to be faster and have lower latency you changed the memory frequency to 500 instead of 400. Or am I reading what you said wrong?

Exactly.
At 400 MHz, your FSB thoughput is 12.8 GB/s.
At 1:1 your memory throughput is 6400 MB/s.
At 4:5 your memory throughput is 8000 MB/s.

It's simple math. 8000 is better than 6400.

[ChaosMinionX]

Exactly.
At 400 MHz, your FSB thoughput is 12.8 GB/s.
At 1:1 your memory throughput is 6400 MB/s.
At 4:5 your memory throughput is 8000 MB/s.

It's simple math. 8000 is better than 6400.

Yes, but of course running a higher memory is going to be faster regardless of 1:1 or 4:5, but 400fsb 1:1 DDR2 800mhz on 533 is going to run faster than say 319fsb 4:5 DDR2 800mhz on 667 mem frequency.

So Its not necessarily an ideal comparison just that your memory is being overclocked by doing so, so of course it would be faster. Or am I interpreting your statement wrong gtj?


EDIT: However to Contribute to your statement regarding bandwidth..... heres a quick bench I did.

FSB: 400mhz / ref freq. 266 / mem ref 533 DDR2 800 1:1 3-3-3-6

--------[ EVEREST Ultimate Edition 2007 (c) 2003-2007 Lavalys, Inc. ]---------------------------------------------------

Version EVEREST v3.80.951 Beta
Benchmark Module 2.1.184.0
Homepage http://www.lavalys.com/
Report Type Quick Report
Computer DEUSEXMACHINA
Generator AjaX
Operating System Microsoft Windows XP Professional 5.1.2600 (WinXP Retail)
Date 2007-04-03
Time 23:28


--------[ Memory Read ]-------------------------------------------------------------------------------------------------

Xeon 3060 3600 MHz Intel Bad Axe 2 D975XBX2 i975X Dual DDR2-800 3-3-3-6 9917 MB/s
--------[ Memory Write ]------------------------------------------------------------------------------------------------

Xeon 3060 3600 MHz Intel Bad Axe 2 D975XBX2 i975X Dual DDR2-800 3-3-3-6 7270 MB/s
--------[ Memory Copy ]-------------------------------------------------------------------------------------------------

Xeon 3060 3600 MHz Intel Bad Axe 2 D975XBX2 i975X Dual DDR2-800 3-3-3-6 7210 MB/s
--------[ Memory Latency ]----------------------------------------------------------------------------------------------

Xeon 3060 3600 MHz Intel Bad Axe 2 D975XBX2 i975X Dual DDR2-800 3-3-3-6 43.5 ns



FSB: 400mhz / ref freq 266 / mem ref 667 DDR2 1000 4:5 4-4-4-12

--------[ EVEREST Ultimate Edition 2007 (c) 2003-2007 Lavalys, Inc. ]---------------------------------------------------

Version EVEREST v3.80.951 Beta
Benchmark Module 2.1.184.0
Homepage http://www.lavalys.com/
Report Type Quick Report
Computer DEUSEXMACHINA
Generator AjaX
Operating System Microsoft Windows XP Professional 5.1.2600 (WinXP Retail)
Date 2007-04-03
Time 23:32


--------[ Memory Read ]-------------------------------------------------------------------------------------------------

Xeon 3060 3600 MHz Intel Bad Axe 2 D975XBX2 i975X Dual DDR2-1000 4-4-4-12 10014 MB/s
--------[ Memory Write ]------------------------------------------------------------------------------------------------

Xeon 3060 3600 MHz Intel Bad Axe 2 D975XBX2 i975X Dual DDR2-1000 4-4-4-12 7293 MB/s
--------[ Memory Copy ]-------------------------------------------------------------------------------------------------

Xeon 3060 3600 MHz Intel Bad Axe 2 D975XBX2 i975X Dual DDR2-1000 4-4-4-12 7749 MB/s
--------[ Memory Latency ]----------------------------------------------------------------------------------------------

Xeon 3060 3600 MHz Intel Bad Axe 2 D975XBX2 i975X Dual DDR2-1000 4-4-4-12 44.0 ns


For me DDR800 3-3-3-6 is rock solid on orthos for 24hrs+, with 2.32v
DDR2 1000 wont even boot cas 3 with anything around that voltage, so I went with the usual 4-4-4-12 and thats what I came up with marginal increases if at all between using either 1:1 or 4:5 however the timings of my DDR2 800 is very very tight for 24/7 thats why I paid for the domc3's but if you dont I suppose 4-4-4-12 DDR2 800 would get pounded by DDR2 1000. So I will go ahead and agree with Gtj on his statement

[gtj]

Yes, but of course running a higher memory is going to be faster regardless of 1:1 or 4:5, but 400fsb 1:1 DDR2 800mhz on 533 is going to run faster than say 319fsb 4:5 DDR2 800mhz on 667 mem frequency.

So Its not necessarily an ideal comparison just that your memory is being overclocked by doing so, so of course it would be faster. Or am I interpreting your statement wrong gtj?

Well, I just tested the the 2 scenarios you describe and 400,1:1 did run slightly faster memory throughput than 319,4:5 but I think it's more to do with the processor running at 3.6 GHz instead of 2.88 GHz

Also, in order to do a valid test, you have to run the memory in single channel mode (1 stick) otherwise the effects of running both channels skews the results.

So I will go ahead and agree with Gtj on his statement

[/QUOTE]

Gee thanks.

[Forceman]

This is easy to test. Set your FSB to a value that let's you run 1:1 plus either 2/3 or 4/5 reliably. Run an Everest Cache and Mmeory benchmark at 1:1, then without changing anything else, set a 2/3 or 4/5 and run the benchmarks again. Throughput should go up and latency should go down. What a concept.

If anyone decides to test this, go ahead and run some other tests as well to see what kind of benefit this has in the real world. Maybe some game benchmarks or SuperPi (actually, that's probably not memory intensive) - anyway something to verify that it is better to run asynch.