8SQ800 vs. 4PEA+ Test Results

[deeznuts]

Well, I finally got the Gigabyte up and running. It's the 8SQ800 from newegg, non-ultra, without digital sound, raid, or SATA. The board itself looks nice, has a good layout, and has the usual gigabyte build quality. The northbridge is A0 stepping, not the newer B0, and revision is 1.0.

Issues I have had with the board:
This board, at least the one I have, is not too stable. It installs XP perfectly everytime, no problem. But once XP is installed, it is finnicky, and about 1 in every 6 reboots, refuses to load windows. Whenever it fails to load Windows, hitting the reset button will not work, as it hangs when detecting the IDE drives at post. A hard reboot (shutting down, then turning it back on) cures this, and it goes into windows.
Another issue, I don't know how major, is that when running Sandra tests, in the summary below the results, Sandra always says that the SMP test has run, and I'm only allowed to choose between SMP results. I think this is related to the Hyper THreading workaround that Gigabyte implemented into the A0 northbridge, which was not qualified to run Hyper Threading. I don't know if this is what's causing Windows to fail to load every once in a while, as the board is solid once it gets there.
Another issue is Vdimm, which only goes to 2.6. Because of this, I couldn't run 4:5 ratio when overclocking my 2.4B to 167 fsb. And even at DDR400, i had to back down the timings.
Another issue is for some reason, Windows takes forever to load, like it keeps hesitating when it should be flying through. the epox is much faster in this department.
The last major issue is resetting the CMOS. What a pain in the butt. There is no jumper, you must physically take out the battery for 30 seconds or so. I use a SCSI card, and must remove it just to get to the battery.

Test Setup (all tests done on SiS655 was dual channel)

Variables Common to Both Platforms - 2.4B C1 Pentium 4, Radeon 9700 (non-pro) at default speeds, Turtle Beach Santa Cruz, 2 x 256MB Twinmos DDR400 with winbond 5ns chips, Adaptec 2930U2 SCSI controller, 18GB Atlas 10KIII, Seagate Cuda IV 80GB, Lite-On 52x, Antec PP412X. Clean install of Windows XP.

-4PEA+ Specifics - Newest Intel INF's, and Intel Application Accelerator
-8SQ800 Specifics - SiS AGP Driver

Different Configurations
1. i845Pe, 2.4, 4:5 ratio
SPeed - 2411.6
Mem Freq - 334.9
Mem Timings - 2-2-2-5

2. i845PE, 3.0, 4:5 ratio
FSB Bios - 167
FSB CPUZ - 167
Speed - 3006.8
Mem Freq - 417.6
Mem Timings - 2-2-2-5

3. SiS655, 2.4, 1:1 ratio
FSB Bios - 133
FSB CPUZ - 133.9
Speed - 2410.2
Mem Freq - 266 (dual channel)
Mem Timings - 2-2-2-5

4. SiS655, 2.4, 4:5 ratio
FSB Bios - 133
FSB CPUZ - 133.9
Speed - 2410.2
Mem Freq - 333
Mem Timings - 2-2-2-5

5. SiS655, 3.0, 1:1 ratio
FSB Bios - 167
FSB CPUZ - 167
Speed - 3006.8
Mem Freq - 417.6
Mem Timings - 2-2-2-5

6. SiS655, 3.0, DDR400 (couldn't quite reach 4:5 ratio)
FSB Bios - 167
FSB CPUZ - 167
Speed 3006
Mem Frequ - 400
Mem Timings 2.5-3-3-6

Tests Ran - Sandra 2003 Mem bandwidth, Arithmetic, and Multimedia. 3dmark2001 (build 330, everything default), and TMPGEnc encoding of 9000 frames (about 5 minutes) at 2500 bitrate.
Not an exhaustive suite of benchmarks, I know, but I really don't have access to any more. That's all I have! I tried to run 3dmark2003, but the damn thing takes to long, so in the interests of my time, I stopped testing it.

Sandra 2003 Mem Bandwidth
Int 3820, Float 3818 - SiS655, 3.0, DDR400
Int 3770, Float 3772 - SiS655, 3.0, 1:1 ratio
Int 3311, Float 3308 - SiS655, 2.4, 4:5 ratio
Int 3189, Float 3187 - i845PE, 3.0, 4:5 ratio
Int 3090, FLoat 3094 - SiS655, 2.4, 1:1 ratio
Int 2550, Float 2548 - i845PE, 2.4, 4:5 ratio

Sandra 2003 Arithmetic
Dhrystone 8245, Whetstone 1716/3925 - SiS655, 3.0, 1:1 ratio
Dhrystone 8217, Whetstone 1716/3923 - SiS655, 3.0, DDR400
Dhrystone 7690, Whetstone 1741/3958 - i845Pe, 3.0, 4:5 ratio
Dhrystone 6718, Whetstone 1393/3171 - i845PE, 2.4, 4:5 ratio
Dhrystone 6705, Whetstone 1377/3150 - SiS655, 2.4, 4:5 ratio
Dhrystone 6689, Whetstone 1374/3139 - SiS655, 2.4, 1:1 ratio

Sandra 2003 MultiMedia
Int 11943, Float 15212 - i845PE, 3.0, 4:5 ratio
Int 11843, Float 15097 - SiS655, 3.0, 1:1 ratio
Int 11842, Float 15091 - SiS655, 3.0, DDR400
Int 9570, Float 12197 - i845PE, 2.4, 4:5 ratio
Int 9500, Float 12120 - SiS655, 2.4, 4:5 ratio
Int 9500, Float 12115 - SiS655, 2.4, 1:1 ratio

3dMark2001 Build 330
14356 - SiS655, 3.0, DDR400
14319 - i845PE, 3.0, 4:5 ratio
14315 - SiS655, 3.0, 1:1 ratio
13413 - SiS655, 2.4, 4:5 ratio
13091 - i845PE, 2.4, 4:5 ratio
13084 - SiS655, 2.4, 1:1 ratio

TMPGEnc - 9000 frames, ~5min, 2500 bitrate
5:06 - i845PE, 3.0, 4:5 ratio
5:12 - Sis655, 3.0, DDR400
5:13 - SiS655, 3.0, 1:1 ratio
6:20 - SiS655, 2.4, 4:5 ratio
6:26 - SiS655, 2.4, 1:1 ratio
6:35 - i845PE, 2.4, 4:5 ratio

Again, this is not meant to be an exhaustive test, nor even a perfect one. Are my scores off? I don't know, I don't benchmark often, so I don't know tweaks, or quirks, etc. Most of the reviews on the net show the SiS performing well when using the memory dividers, and not a 1:1 ratio. I still don't know why the dividers help, when theoretically they shouldn't, but they do. That being said, the i845PE performed quite well when overclocked at 4:5 ratio was held. Of course, if I could get my ram to run 4:5 on the sis, when overclocked, and at good timings, it should have increased scores. But a 3:4 ratio on the i845PE might still be there with them.

Anyways, that's my little tiny review, don't take everything literal. And let me know if you see mistakes.

 

[Yozza]

Wow! Thanks for taking the time and effort to post these results.

I'm surprised how well your i845PE keeps up with the SIS655 at 167MHz FSB, though the tighter memory timings could be a contributing factor to that.

Could you do a quick memory test with ScienceMark 2.0? That'll determine the latency of SIS655's memory controller vs the i845PE, and do possibly also Sandra's unbuffered memory benchmark (the one with 9 boxes unticked - all SIMD instructions disabled, and the box near the bottom for "buffering" unticked too). The unbuffered benchmark should show how much effective bandwidth is available, rather than the absolute (and unrealistic) maximum obtained from the buffered benchmark.

Again, thanks for taking the time to share your experiences with us

 

[bluesky]

Thanks.
It seems it's not the board I want to get.

 

[deeznuts]

ok, here are the science mark results. no 4pEA+ because i already took that out hehe.

2.4 1:1 ratio
bandwidth 2550.83 mb/s
5 cycles / 2.07 ns /4 byte stride
20 cycles / 8.30 ns / 16 bit stride
74 cycles / 30.70 ns / 64 bit stride
301 cycles / 124.88 ns / 256 byte stride
306 cycles / 126.96 ns / 512 bye stride

2.4 4:5 ratio
bandwidth - 2746.38 mb/s
5 - 2.07 ns / 4 byte stride
20 - 7.47 ns / 16 byte stride
65 - 26.96 ns / 64 byte stride
259 - 107.42 ns / 256 byte stride
260 - 107.83 ns / 512 byte stride

3.0 1:1 ratio
bandwidth 3153.62 mb/s
5 - 1.67 ns / 4 byte
20 - 6.67 ns / 16 byte
76 - 25.33 ns / 64 byte
316 - 105.32 ns / 256 byte
321 106.99 ns / 512 byte

3.0 DDR400 (relaxed timings)
bandwidth 3232.63 mb/s
5 - 1.67 ns / 4 byte
20 - 6.67 ns / 16 byte
73 - 99.65 ns / 64 byte
299 - 99.65 ns / 256 byte
304 / 101.31 ns / 512 byte

maybe you can explain what the results mean hehe. still don't understand how to do the unbuffered tests in sandra. i open the option box, and do see the buffered checkmark, but what nine other boxes? there are tons in there.

also, another note is that i just flashed my bios to F4, it shipped with F3, and it does seem more stable. gonna keep rebooting to see if it ever hangs again.

 

[oldfart]

Glad to see the TMPGEnc benches. I get tired of the same old 3DMark/game bench stuff.

 

[deeznuts]

Originally posted by: oldfart
Glad to see the TMPGEnc benches. I get tired of the same old 3DMark/game bench stuff.

yeah, it's the only thing i had laying around that i could use to benchmark, and that i actually know how to use hehe.

 

[jpetermann]

Nice tests!

PJ

 

[deeznuts]

ok, using a modded bios which allows higher vdimm, i was able to run a 4:5 ratio while at 3.0Ghz.

-SiS655 3.0 4:5 ratio
FSB Bios - 167
FSB CPUZ - 166.7
Speed - 3000.7
Mem Freq - 416
Mem Timings - 2-3-3-7
Sandra Mem Bandwidth
Int - 3895 mb/s
Float - 3895 mb/s
Sandra Arithmetic
Dhrystone - 8247
Whetstone - 1715/3925
Sandra Multimedia
Integer - 11845
Floating Point - 15100
3dmark2001
14460

can't do the TMPGEnc test cuz my trial ran out damn!

 

[NFS4]

Originally posted by: deeznuts
ok, using a modded bios which allows higher vdimm, i was able to run a 4:5 ratio while at 3.0Ghz.

-SiS655 3.0 4:5 ratio
FSB Bios - 167
FSB CPUZ - 166.7
Speed - 3000.7
Mem Freq - 416
Mem Timings - 2-3-3-7
Sandra Mem Bandwidth
Int - 3895 mb/s
Float - 3895 mb/s
Sandra Arithmetic
Dhrystone - 8247
Whetstone - 1715/3925
Sandra Multimedia
Integer - 11845
Floating Point - 15100
3dmark2001
14460

can't do the TMPGEnc test cuz my trial ran out damn!

1) Where'd you get the modded BIOS and who made it??
2) Can you upgrade to the modded BIOS from within Windows using Gigabyte's @BIOS program?
3) How far will it take the DIMM voltage up to? I'm currently running at the max 2.6 on my 8SQ800
4) MUST HAVE THE BIOS FILE, email it if you can please
5) Also, is the modded BIOS based on the F4 version?

 

[Yozza]

Originally posted by: deeznuts
ok, here are the science mark results. no 4pEA+ because i already took that out hehe.
maybe you can explain what the results mean hehe. still don't understand how to do the unbuffered tests in sandra. i open the option box, and do see the buffered checkmark, but what nine other boxes? there are tons in there.

also, another note is that i just flashed my bios to F4, it shipped with F3, and it does seem more stable. gonna keep rebooting to see if it ever hangs again.

Thanks for the ScienceMark results. They are quite interesting, and show that the SIS655 has a much higher memory latency than E7205 (Granite Bay) and i845PE. ~300 cycles is in RDRAM territory! For comparison, Granite Bay achieves ~80ns latency at 167MHz FSB.

However, what it does show is that increasing the memory frequency and running it asynchronously does help considerably with the memory latencies. This is interesting, as some predicted latencies would stay roughly the same or would increase because of the buffering needed to achieve asynchronous operation. Well, you can clearly see that the increased clock granularity greatly offsets this async latency .

For unbuffered SANDRA Memory benchmarks, untick the boxes to do with SIMD instructions - MMX, SSE, SSE2, etc. There are 8 boxes for that, and then 1 for "Enable buffering/pre-fetech benchmarks".


NFS4, see this thread at AsusBoards for BIOS download links. It allows many previously hidden options, as well as DIMM voltage adjustments upto 2.8V!

 

[deeznuts]

so that's how you do unbuffered tests! hehe, could never figure it out. thanks.

about the latency, wow, never knew it was that high. maybe that explains why it doesn't do quite as well until you change the ratios up. this doesn't help much if you are overclocking like crazy with a high fsb. in my tests, it seems at 3.0Ghz, the PE with 4:5 ratio beats up on, or at least hangs with, the dual channel when at 1:1.

i'm figuring a lot of people are not pleased with my results, as I wasn't either. i really wanted the SiS to kick some butt, but it did not, at least in my situation. it might be going back.