45nm / DDR3 era Spi discuz.

[VictorWang]

after playing for months ddr3/45nm
i do feel that less and less effective on tweak a 45nm/ddr3 system.
dont know if someone have the same feeling as I do.

here is 445*9=4G pi_1m/32m & 450*10=4.5G pi_1m/32m results.

maxmem=624m, largesystemcache=1, non-pagefile, copy waza 1.36G.


1) 445*9=4G 1m = 11.468sec


2) 445*9=4G 32m=11min00.562sec



'coz of FSB limited, so it seems damn hard to break 11min
---------------------------------------------------------------------

3) 450*10 1m=10.219sec (can break 10.1sec ?)



4) 450*10 32m=9min59.531sec (just break 10min, any possibility for 9min55sec ? )

[Zeus]

Victor,

I cannot speak from own experience cuz i don't have a 45nm cpu nor do i have DDR3.
From observation it looks like the large cache on the 45nm cpu's is responsible for the great performance in SuperPi but the downside is that tweaks like copy-waza take less effect.

Looking back at A64 days with 512mb cache CPU's, copy-waza gave some impressive gains, even in 1M, which reinforces the feeling that more cache means less gain from CW.

For DDR3, it looks like this is only taking a lead over DDR2 when ran at outrages speeds and relative tight timings.
Running more relaxed than 6-7-7 on DDR3 does not give any gain over DDR2 imo.

[Gautam]

Simply put, the effect is the same as what we saw from going from A64 to Conroe. SuperPi is much less tweakable, since most of our tweaks deal with memory. With a larger cache, there is much less memory access.

[BeardyMan]

compare with DDR2 woudl be cool

example at
600 4_4_4 you need DDR3 at etc etc

[dinos22]

yeah i've played a little and found similar things

i have to say that bioses are still in early stages which will improve performance once they become more stable

i've mainly played with GB boards and the dividers are pretty spotty and hard to tweak but when you hit the right ratio speeds are REALLY good >> 45K 1M efficiency non tweaked with memory not maxed out yet at 5GHz....

i am looking forward to another chipset manufacturer introducing their chipsets which will have dividers that can allow RAM to be maxed out at stock FSB as well as any other frequency. That is one thing that bugs me about Intel chipsets.....to take advantage of the new RAM you have to hit high FSB and with quads on the way most people will be maxing out around 450 FSB which is just not enough

FSB plays a big role in SuperPi so we really need those E8500 chips to start hitting 600 to see where we're at and whether it's the FSB as well that is lacking

Victor is right though tweaking isn't as effective....i tried mainly 1M stuff but i see 32M is definitely not as good as it used to be with lower cache CPU from you and hipro

[erickwok]

after playing for months ddr3/45nm
i do feel that less and less effective on tweak a 45nm/ddr3 system.
dont know if someone have the same feeling as I do.

here is 445*9=4G pi_1m/32m & 450*10=4.5G pi_1m/32m results.

maxmem=624m, largesystemcache=1, non-pagefile, copy waza 1.36G.


1) 445*9=4G 1m = 11.468sec


2) 445*9=4G 32m=11min00.562sec



'coz of FSB limited, so it seems damn hard to break 11min
---------------------------------------------------------------------

3) 450*10 1m=10.219sec (can break 10.1sec ?)



4) 450*10 32m=9min59.531sec (just break 10min, any possibility for 9min55sec ? )

Wa! Super Fast!
whats motherboard?
whats ram? and voltage??

[VictorWang]

Wa! Super Fast!
whats motherboard?
whats ram? and voltage??

Asus Max Extreme.
Cellshocks DDR3 1800, 2.2v

yeah i've played a little and found similar things

i have to say that bioses are still in early stages which will improve performance once they become more stable

i've mainly played with GB boards and the dividers are pretty spotty and hard to tweak but when you hit the right ratio speeds are REALLY good >> 45K 1M efficiency non tweaked with memory not maxed out yet at 5GHz....

i am looking forward to another chipset manufacturer introducing their chipsets which will have dividers that can allow RAM to be maxed out at stock FSB as well as any other frequency. That is one thing that bugs me about Intel chipsets.....to take advantage of the new RAM you have to hit high FSB and with quads on the way most people will be maxing out around 450 FSB which is just not enough

FSB plays a big role in SuperPi so we really need those E8500 chips to start hitting 600 to see where we're at and whether it's the FSB as well that is lacking

Victor is right though tweaking isn't as effective....i tried mainly 1M stuff but i see 32M is definitely not as good as it used to be with lower cache CPU from you and hipro

ratio is a problem.
fsb is a problem especially on 45nm both wolf/yorkies
and bios/spd seems also have some problems
maxmem gives only about 1 or 2sec improvement
copzy waza only about 2sec or 3sec.

Simply put, the effect is the same as what we saw from going from A64 to Conroe. SuperPi is much less tweakable, since most of our tweaks deal with memory. With a larger cache, there is much less memory access.

we might need to tweak "write" peformance while as copy/cache tweak "Read".
so a Good HDD or I-Ram is needed.
less time and cpu usage when pi_32m "write" data from Ram to HDD might be give another 2 or 3sec faster.

Victor,

I cannot speak from own experience cuz i don't have a 45nm cpu nor do i have DDR3.
From observation it looks like the large cache on the 45nm cpu's is responsible for the great performance in SuperPi but the downside is that tweaks like copy-waza take less effect.

Looking back at A64 days with 512mb cache CPU's, copy-waza gave some impressive gains, even in 1M, which reinforces the feeling that more cache means less gain from CW.

For DDR3, it looks like this is only taking a lead over DDR2 when ran at outrages speeds and relative tight timings.
Running more relaxed than 6-7-7 on DDR3 does not give any gain over DDR2 imo.

i tried lots of different mem_timing combination.
find the tighten tRP and tRAS will help nth but results in un-stable and might be slower.
but tweak sub_timings usually gives a boost (about 2sec).

[mrlobber]

we might need to tweak "write" peformance while as copy/cache tweak "Read".
so a Good HDD or I-Ram is needed.
less time and cpu usage when pi_32m "write" data from Ram to HDD might be give another 2 or 3sec faster.

I-RAM didn't help in Superpi 32M with Conroe already What makes you think with Yorkfield/Wolfdale this would be different?

[massman]

These quite fit the thread. Results based on 265 45nm results:

[dinos22]

massman so there are now a fair few sub 45K runs

who are they from

[mrlobber]

These quite fit the thread. Results based on 265 45nm results:

Hoho, quite a nice curve You might want to remove that outlier though (or is it included with an intent since it falls out of all the other pack not by 100%, but by 1000% at least ) and might impact the equation of the curve quite significantly.

[Gautam]

How are you getting those graphs massman? Or rather, the data for them.

[BeardyMan]

How are you getting those graphs massman? Or rather, the data for them.

LOL
i guess hwbot my friend

[Gautam]

Yeah, I guessed hwbot too. But is there a way to extract all those numbers?

[metro.cl]

Massman in that graph DDR2 and DDR3 is mixed? you could color DDR2 diferently?

[T_M]

You might want to remove that outlier though

Sconded, then there'd be greater visibility of the rest

[massman]

massman so there are now a fair few sub 45K runs

who are they from

Kevin. I have seen very low 45k runs as well with low memory clocks, but I still have to contact that person to ask how exactly he got those numbers.

Hoho, quite a nice curve You might want to remove that outlier though (or is it included with an intent since it falls out of all the other pack not by 100%, but by 1000% at least ) and might impact the equation of the curve quite significantly.

I included the 1.2G run to proof that the 'prediction line' (please, what's the correct term in English?) is still quite accurate at very slow speeds.

How are you getting those graphs massman? Or rather, the data for them.

HWBot, Ripping.org, XS, OCX, Coolaler boards

Yeah, I guessed hwbot too. But is there a way to extract all those numbers?

No, manually added in my excel chart.

Massman in that graph DDR2 and DDR3 is mixed? you could color DDR2 diferently?

I'll try, my excel skills are not that good

[VictorWang]

great job massman.

have a 5.4G 32m result, not bad

[massman]

DDR2 and DDR3 split up (less data as I don't know for every result what kind of ram was used). Can't make any conclusion, but I suspect that low latencies can outperform high bandwidth.

[KTE]

3 of the DDR2 runs and 2 of the DDR3 runs are a cut above the rest. By quite a large margin for something that simple, quick and so often run.

Thanks for the charts. How long do you spend on such things massman?
Man you guys have some time or what. Whose got the furthest left DDR2 run?

[massman]

Long enough to realise that I do not have a 45nm chip to test my theories with .

Furthest left: Rolco 11,265 - 4001 - 600 - 45071,265

[KTE]

Thanks mate.

[VictorWang]

45nm 4Ghz 1m < 11.4sec is damn damn fast.
any idea how to achieve that ?

[Rol-Co]

45nm 4Ghz 1m < 11.4sec is damn damn fast.
any idea how to achieve that ?

Yep..... 11.234 @4g now

[T_M]

Whats ur secret ? :P