Real-world effect of i975 1333 FSB strap vs. 1066 strap (and more Conroe performance)

[uOpt]

In this thread:
http://www.xtremesystems.org/forums/...77&postcount=1
Tony explains that choosing the higher 1333 FSB strap might lead to
higher clocks, but that latency is pessimized. I had previously
observed that memory performance stuffers


Here are some real numbers. Basically it compares a E6600 at 1333 *
120% = 3660 MHz to the same setup at 1066 * 150% = 3600 MHz. The
memory in question is the Supertalent DDR2-1000. I had to take it to
800 MHz with the 1066 strap but as you see even then it comes out
faster.

Results (short ones best of three):

                                Conroe  1066*1.50       idx1    idx2
FC5/amd64:                      1333*1.20      Opt175@2.6
---------------------------------------------------------------------------
stream MB/sec                 3951    4411     2469     1.116   0.62
drystones Mio                 16.7    16.5     7.5      0.988   0.45
superpi 1 M sec               10.6    10.5     26.4     0.991   2.49
superpi 4 M sec               -1      58.1     141.3
superpi 32 M sec              824     782      1755     0.949   2.13
dual superpi 32 M sec avg     1066    1007     2000     0.945   1.88
Linux -j 32 real              487     459      713      0.943   1.46
Linux -j 32 cpu               868     853      1395     0.983   1.61
Watts idle                    178     180(175) 120      1.011   0.67
Watts Linux -j 32             247     251(243) 217      1.016   0.88
Watts dprime 5 min            279     285(275) 230      1.022   0.82
FreeBSD 32 bits make world    1114    1126     1749     1.011   1.57
Linux -j 32 real (in FB)      132     133      202      1.008   1.53
Linux -j 32 cpu (in FB)       247     244      390      0.988   1.58

"idx1" and "idx2" are the relative values of the 1066*1.50 run and the
Opteron run against the 1333*1.20 run.

Watts in (...) were at 1.450 V, real run at 1.475 V. Note that both
Conroe runs are at 1.475 Vcore and that the 1066*1.50 has lower Vdimm
but higher Vmch. Watts for Conroe include waterpump which takes about
20 watts, Opteron is on Freezer 64 Pro. Conroe has 2x 1 GB,
Opteron939 has 4x 1 GB ECC.

The latter Linux kernel in FB tests are building a small 32bit Linux
kernel using a 32 bit Linux compiler in the FreeBSD Linux compat
environment.

As you can see, this 32 bit environment, even in the Linux layer of
FreeBSD, reacts precisely to the CPU in use like the Linux native 64
bits. This rules out that somehow magically Conroe is weak at 64 bits
and it also rules out that somehow FreeBSD's Linux emulator is somehow
magically screwing something up.

However, there is tendency that the 1066*1.50 run, which has lower
clockspeed but higher memory bandwidth, on average likes the 64 bit
code more. Hardly surprising and the difference is too small to make
a fuss out of it.

I'll followup with the exact settings in a minute.

[uOpt]

1) High FSB strap settings:
1333 * 1.20 = 3660 MHz, 1.475 Vcore
667-5-4-4-12 = 1000 MHz, 2.1 Vdimm 3951 MB/sec stream (vmch/vfsb = default)

2) Lower base FSB, lower latency:
1066 * 1.50 = 3600 MHz, 1.475 Vcore
533-4-4-4-11 = 800 MHz, 1.9 Vdimm 1.600Vmch/ def Vfsb/101MHz-PCIe

3) DC Opteron at 2.60 GHz @ 1.60V, memory 416 MHz 3-3-3-8 2T PCI hole
remapping active

My benchmark suite's opinion on all this:

User CPU speed relative to Core2 3.00 GHz (Conroe x2), 2x 666(?) MHz 2048 MB, i975 Bad Axe (SMP kernel) [2 CPUs]

Opteron 3.08 GHz (148 skt 939), 2x 342 MHz 1T 3-4-4-8 TCCD Geil One 512 MB, NF4 SLI-DRI (SMP kernel) [1 CPUs]:
            Normal C compilation:   90.7 ( 105.9)
                 C++ compilation:   92.8 ( 108.3)
               Linux compilation:   89.2 ( 104.2)
                            gzip:   84.5 (  98.6)
                            Lisp:   80.6 (  94.1)
                          python:   76.0 (  88.8)
                          php/bs:   97.3 ( 113.6)
              High quality mpeg4:   87.3 ( 101.9)
          Constant bitrate mjpeg:   79.9 (  93.2)
                      Fast mjpeg:   78.3 (  91.4)
                         Overall:   85.7

Opteron DC 2.60 GHz (175 skt 939), 4x 216 MHz 2T 3-4-4-8 -5B D 4096 MB (3 GB soft), Via A8V-E SE (SMP kernel) [2 CPUs]:
            Normal C compilation:   75.8 ( 105.2)
                 C++ compilation:   76.2 ( 105.7)
               Linux compilation:   73.9 ( 102.6)
                            gzip:   70.8 (  98.3)
                            Lisp:   66.9 (  92.9)
                          python:   64.3 (  89.2)
                          php/bs:   93.7 ( 130.1)
              High quality mpeg4:   72.7 ( 100.9)
          Constant bitrate mjpeg:   63.9 (  88.7)
                      Fast mjpeg:   62.3 (  86.4)
                         Overall:   72.1

Core2 2.40 GHz (Conroe x2), 2x 533 MHz 2048 MB, i975 Bad Axe (SMP kernel) [2 CPUs]:
            Normal C compilation:   79.8 (  99.1)
                 C++ compilation:   79.7 (  99.0)
               Linux compilation:   79.8 (  99.1)
                            gzip:   78.8 (  97.8)
                            Lisp:   79.3 (  98.5)
                          python:   79.6 (  98.9)
                          php/bs:   89.0 ( 110.5)
              High quality mpeg4:   79.8 (  99.1)
          Constant bitrate mjpeg:   79.7 (  99.0)
                      Fast mjpeg:   79.7 (  99.0)
                         Overall:   80.5

Core2 3.00 GHz (Conroe x2), 2x 666(?) MHz 2048 MB, i975 Bad Axe (SMP kernel) [2 CPUs]:
            Normal C compilation:  100.0 ( 100.0)
                 C++ compilation:  100.0 ( 100.0)
               Linux compilation:  100.0 ( 100.0)
                            gzip:  100.0 ( 100.0)
                            Lisp:  100.0 ( 100.0)
                          python:  100.0 ( 100.0)
                          php/bs:  100.0 ( 100.0)
              High quality mpeg4:  100.0 ( 100.0)
          Constant bitrate mjpeg:  100.0 ( 100.0)
                      Fast mjpeg:  100.0 ( 100.0)
                         Overall:  100.0

Core2 3.60 GHz (Conroe x2), 2x 800 MHz 1T 4-4-4-11 2048 MB (1066 FSB strap), i975 Bad Axe (SMP kernel) [2 CPUs]:
            Normal C compilation:  120.0 ( 101.1)
                 C++ compilation:  120.2 ( 101.3)
               Linux compilation:  120.1 ( 101.2)
                            gzip:  119.6 ( 100.8)
                            Lisp:  119.9 ( 101.0)
                          python:  120.3 ( 101.4)
                          php/bs:  106.9 (  90.1)
              High quality mpeg4:  120.1 ( 101.2)
          Constant bitrate mjpeg:  119.6 ( 100.8)
                      Fast mjpeg:  120.1 ( 101.2)
                         Overall:  118.7

Core2 3.66 GHz (Conroe x2), 2x 1000 MHz 1T 5-4-4-12 2048 MB (1333 FSB strap), i975 Bad Axe (SMP kernel) [2 CPUs]:
            Normal C compilation:  121.2 ( 102.0)
                 C++ compilation:  119.3 ( 100.5)
               Linux compilation:  120.7 ( 101.6)
                            gzip:  121.1 ( 102.0)
                            Lisp:  120.0 ( 101.1)
                          python:  120.5 ( 101.5)
                          php/bs:  106.2 (  89.4)
              High quality mpeg4:  121.1 ( 102.0)
          Constant bitrate mjpeg:  118.8 ( 100.0)
                      Fast mjpeg:  118.8 ( 100.0)
                         Overall:  118.8

[theteamaqua]

interesting

do u know how P5W DH works?? is it all 1066fsb strap?? when u overclock i mean.

[freecableguy]

P5W DH is 1333 strap with no option to change.

[BWR]

Are there any gaming benchmarks?

[uOpt]

P5W DH is 1333 strap with no option to change.

However, I think Tony said that Asus does the 1333 strap without increasing the latency as Intel does.

The latency on the Intel board with 1333 FSB seems to be pretty drastic. 10% less bandwidth with 20% higher RAM clock and 2% higher clockspeed. The 10% less memory bandwidth seem to translate to 2% real time in some real-world applications.

Damn CPUs with low max multipliers

[theteamaqua]

yeah but for 975 XBX i gotta solder (not steady hands :X ) in order to even overclock

[uOpt]

yeah but for 975 XBX i gotta solder (not steady hands :X ) in order to even overclock

No, the Radioshack conductive paint does fine (thanks FCG!).

Also, using the ITK you don't even need to do that.

[Tony]

However, I think Tony said that Asus does the 1333 strap without increasing the latency as Intel does.

The latency on the Intel board with 1333 FSB seems to be pretty drastic. 10% less bandwidth with 20% higher RAM clock and 2% higher clockspeed. The 10% less memory bandwidth seem to translate to 2% real time in some real-world applications.

Damn CPUs with low max multipliers

I have the Asus now, im just waiting on a fresh CPU and I will get testing

I will try to see what registers are being set so we know for sure if ithe strap is 1333 or 1066.

[Tony]

I keep saying this as its real important...if you can not stretch to a 6800 get the 6700, you need the high multi on the 1066 strap for the best speed.

[uOpt]

I keep saying this as its real important...if you can not stretch to a 6800 get the 6700, you need the high multi on the 1066 strap for the best speed.

Seconded. Current E6600 have more steam than the 3600 MHz that they are stuck at with the 1066 strap. Andthat 1066 * 1.50 is hard to get stable, althought not as hard as 1333 * 1.20 to get the same 3600 MHz. A higher multiplier is a must-have if you do high end overclocking.

[eva2000]

Seconded. Current E6600 have more steam than the 3600 MHz that they are stuck at with the 1066 strap. Andthat 1066 * 1.50 is hard to get stable, althought not as hard as 1333 * 1.20 to get the same 3600 MHz. A higher multiplier is a must-have if you do high end overclocking.

i think it depends on each boards chipset latencies as it seems even the same model boards have varying max fsb even at 1067fsb strap

my E6600 QPGE cpus have no probs on badaxe rev304 unmodded at 421fsb on 1067fsb strap and max out at 429fsb on 1333fsb strap

some asus p5w dhs have different max fsbs as well..

now i'm wondering if it's how and what ram is used and how it interacts with chipset latencies i.e. using my 2x512MB OCZ PC5400 Plat XTC on 1067fsb strap max is 378fsb, using 2x512MB 5400ULv1.5 on 1067fsb strap max is 421fsb.

maybe lower latency capable ddr2 modules can handle the more aggressive latencies that come with 1067fsb straps ? it would also explain why asus p5w dh can clock fsb higher if you use 4:3 down clock divider than 1:1

[Charloz24]

Seconded. Current E6600 have more steam than the 3600 MHz that they are stuck at with the 1066 strap. Andthat 1066 * 1.50 is hard to get stable, althought not as hard as 1333 * 1.20 to get the same 3600 MHz. A higher multiplier is a must-have if you do high end overclocking.

Damn......are you saying that my 3.8-4.0 ghz (watercooled) goal with a E6600 is pretty much a no go?

BTW great work on all the benchmarks on differents config!

[uOpt]

Well, one problem is that the BadAxe BIOS overclock shuts off at 150%, which is 400 MHz FSB and results in 3600 MHz CPU with the x9 multiplier on the E6600. On the 1333 strap I get 431 MHz or somesuch but it is harder to get stable at the same effective CPU clockspeed.

Overall the region between 3600 and 3900 is negotiable for the CPU tested in my board but it's too much trouble to worry about it.

I can bench at 1333 * 1.30 = 3900 MHz, but it requires all unmodded volts at full throttle and the machine takes a solid 338 watts out of the A/C dual-mpriming and reported CPU temps reach 86 C quickly, which is 65 to 70 C real. That is on water/storm/dual-rad/jadada.

For comparision, an 8-core 4x Opteron 875 box takes 455 Watts total all cores mpriming with the same harddrive and a comparable PSU.

[uOpt]

Damn......are you saying that my 3.8-4.0 ghz (watercooled) goal with a E6600 is pretty much a no go?

You will reach it but it's a bi1ch to handle, as I just mentioned. The 3600 MHz region is nicely reachable, the 3900 MHz region is going out of bounds power-wise with this CPU. If you have normal ambient temps in your room and good watercooling you will get it benchable, but I have 30-33 C ambient here and non-chilled watercooling.

Possibly, with a more than x9 CPU multiplier the situation is better at 3900 MHz even for the same die quality of the CPU since you can keep FSB, MCH and all that other junk down, both Volt-wise and clock-wise.

[eva2000]

yeah with my OCZ DDR2 ram that does 420mhz 4-4-4-15 easy it didn't boot at 1067fsb strap + 50%, but i stable with Corsair PC5400ULv1.5 with 150% all the way to 421fsb the last 21fsb via clockgen

with 5400ULv1.5 DDR2 can boot into windows easy at 1067fsb + 50%

this was even before i changed NB heatsinks on my badaxe http://www.xtremesystems.org/forums/...3&postcount=64

changing NB heatsinks didn't gain me any fsb at 1067fsb strap but gained 2-4fsb at 1333fsbs trap up from 425fsb to 429fsb

[uOpt]

yeah with my OCZ DDR2 ram that does 420mhz 4-4-4-15 easy it didn't boot at 1067fsb strap + 50%, but i stable with Corsair PC5400ULv1.5 with 150% all the way to 421fsb the last 21fsb via clockgen

I also noticed that my SuperTalent DDR2-1000 could not be operated at 1000 MHz (slowest timings) with the strap at 1066 although it works at 1018 MHz with the strap at 1333.

Still, the 800 MHz RAM with the 1066 strap are considerably faster overall than 1000 MHz RAM with the 1333. The reason is the increased northbridge latency at the 1333 strap which slacks off delivering the otherwise fast RAM results to the CPU. Another reason to demand integrated memory controllers.

I should probably re-run all the above tests on an Asus board with the suspected latency fixed 1333 strap.

[graham_h]

keep us posted with the Asus results Tony