dionb wrote on 2020-05-08, 22:27:

Great stuff 😀
particularly as I own the two fastest boards on the shootout 😜

Have you seen enough Tualatin chipsets for the rest of eternity, or would you be looking to add more tests to this? I'm intrigued where the real strengths of say the Serverworks chipset lie. Not in Q3A, obviously...

My guess for Serverworks: CPU speed and memory bandwidth synthetic benchmarks .

Awesome job! And yeah, VIA chipsets really, really depend on the board's BIOS implementation. They can be really slow, and they can also be faster than 440BX.

This makes me want to rip the 6800GT out of the Barton 3200+ machine that I've been toying with lately and put it back in my PIII-S. The last time I had the 6800GT in the P3, it was "only" running at 150MHz FSB with 2-2-2-6 memory. I've since found PC3200 memory that can tolerate 155MHz 2-2-2-5 operation, and that extra 5MHz of bus speed seems to make a difference! Next COVID-19 project, I guess.

dionb wrote on 2020-05-08, 22:27:

I'm intrigued where the real strengths of say the Serverworks chipset lie. Not in Q3A, obviously...

It's clearly not the ultimate gaming experience, unless your tendencies lie somewhere in the SM-spectrum :d.

The Supermicro is my standard SCSI and ATA disk test bed, and runs very smoothly under Win2K. Under normal operation, I have it installed with two 1,4S, 2*1GB of ECC buffered SDR-SDRAM, an AMCC PCI-X 8-port SATA-II controller and a tiny 40GB Intel SSD. For graphics, a Matrox G450 suffices.

As an all-purpose workstation, it can still holds its own. It's one of those systems you instantly like, because it just feels fast. I suppose the PCI-X helps, and the chipset probably does a better job of of balancing the load onto the two processors than other dual PIII solutions, which I felt to have some irritable lag.

dionb wrote on 2020-05-08, 22:27:

Have you seen enough Tualatin chipsets for the rest of eternity, or would you be looking to add more tests to this?

The hardware's there, so why not. With the common SSD (three partitions, one instance of Win2K per chipsetvendor), it's easy enough to switch from one board to the other. I'm open to suggestions:).

@Standard Def Steve Offline
Q3A (and all games from that time i can think of) when running on a P3 hardware with AGP video cards up to GF 6800# (last AGP compatible ?) is CPU limited.
There is no difference if you use GF/Quadro 2, GF6800U/Quadro FX 4000, or anything in-between really.
At least that's what i see here.
For example: 197 fps roll with Quadro2 Pro https://www.petershipkov.com/temp/retro_pc_im … est_run_150.mp4
With Quadro FX 4000 / GF 6800 Ultra the result is a frame or two higher - ~198-199 fps.

pshipkov wrote on 2021-02-09, 20:00:

@Standard Def Steve Offline Q3A (and all games from that time i can think of) when running on a P3 hardware with AGP video cards […]
Show full quote

@Standard Def Steve Offline
Q3A (and all games from that time i can think of) when running on a P3 hardware with AGP video cards up to GF 6800# (last AGP compatible ?) is CPU limited.
There is no difference if you use GF/Quadro 2, GF6800U/Quadro FX 4000, or anything in-between really.
At least that's what i see here.
For example: 197 fps roll with Quadro2 Pro https://www.petershipkov.com/temp/retro_pc_im … est_run_150.mp4
With Quadro FX 4000 / GF 6800 Ultra the result is a frame or two higher - ~198-199 fps.

Yeah, Quake III is pretty CPU limited. Funnily enough, my 9800 Pro actually managed to beat the 6800 GT by 4 fps in Q3A. Guess NV's superior OpenGL goes out the window when you're completely held back by the CPU. 😜

However, I don't actually play Q3A anymore, and this machine has to handle all of my 2000-2005 PC games. At 1600x1200, quite a few games from 2003-2005 tend to be bottlenecked by the 9800 Pro. The 6800GT, IMHO, is absolutely perfect for this machine. 😀

+4fps !
Moving the needle over there, eh ? 😁

1.63ghz is awesome.
What are you running this on ?
800 something, sis ?

pshipkov wrote on 2021-02-10, 06:28:

+4fps ! Moving the needle over there, eh ? :D […]
Show full quote

+4fps !
Moving the needle over there, eh ? 😁

1.63ghz is awesome.
What are you running this on ?
800 something, sis ?

Apollo Pro 266T, on the stupid-fast QDI Advance 12T motherboard. Seriously, of all the PIII boards I've played with, this one is easily the fastest. Clock-for-clock, it's around 8% faster than a Gigabyte board based on the same chipset, and it smokes my 440BX, 815, and 694T boards as well.

Standard Def Steve wrote on 2021-02-10, 19:25:

pshipkov wrote on 2021-02-10, 06:28:

+4fps ! Moving the needle over there, eh ? :D […]
Show full quote

+4fps !
Moving the needle over there, eh ? :D

1.63ghz is awesome.
What are you running this on ?
800 something, sis ?

Apollo Pro 266T, on the stupid-fast QDI Advance 12T motherboard. Seriously, of all the PIII boards I've played with, this one is easily the fastest. Clock-for-clock, it's around 8% faster than a Gigabyte board based on the same chipset, and it smokes my 440BX, 815, and 694T boards as well.

Isn't that a single-CPU motherboard? I've been pretty impressed with my [dual] boards based on the Apollo Pro 266T chipset. If you can find one which overclocks, has Tualatin support (either via natively or using the Korean adapted CPUs), and AGP 4X, that is they way to go. You do loose out on 64-bit PCI and dual channel SDRAM. I am using one such board from MSI which I run at 1.5 GHz. Although I can set the FSB higher, I didn't see any adjustment for the CPU's core voltage, so I'm guessing why that is why the max it will run at is 1.55 GHz.

feipoa wrote on 2021-02-10, 21:22:

Standard Def Steve wrote on 2021-02-10, 19:25:

pshipkov wrote on 2021-02-10, 06:28:

+4fps ! Moving the needle over there, eh ? :D […]
Show full quote

+4fps !
Moving the needle over there, eh ? 😁

1.63ghz is awesome.
What are you running this on ?
800 something, sis ?

Apollo Pro 266T, on the stupid-fast QDI Advance 12T motherboard. Seriously, of all the PIII boards I've played with, this one is easily the fastest. Clock-for-clock, it's around 8% faster than a Gigabyte board based on the same chipset, and it smokes my 440BX, 815, and 694T boards as well.

Isn't that a single-CPU motherboard? I've been pretty impressed with my [dual] boards based on the Apollo Pro 266T chipset. If you can find one which overclocks, has Tualatin support (either via natively or using the Korean adapted CPUs), and AGP 4X, that is they way to go. You do loose out on 64-bit PCI and dual channel SDRAM. I am using one such board from MSI which I run at 1.5 GHz. Although I can set the FSB higher, I didn't see any adjustment for the CPU's core voltage, so I'm guessing why that is why the max it will run at is 1.55 GHz.

Yep, it's a single CPU board. Like you, I could never overclock dual processor boards as high, and since I use my P3 machines primarily for gaming, I've always favoured higher bus clocks over dual processor capability.

I've never owned a motherboard with dual-channel SDR. Pretty neat idea, but wouldn't 128-bit SDR have the same maximum theoretical bandwidth as 64-bit DDR? Even with single channel DDR, I appear to be at the limit of what the P3's FSB can handle (1252 MB/s memory read bandwidth in AIDA64).

When I tested this some years ago, the dual channel SDR did indeed show better memory benchmark results compared to DDR w.r.t. Tualatins at 1.4 GHz.

Bank interleaving is not dual-channel though.

Even with single channel DDR, I appear to be at the limit of what the P3's FSB can handle (1252 MB/s memory read bandwidth in AIDA64).

Technically, increasing RAM speed above FSB limit is still useful. You can decrease latency and improve performance for DMA devices like AGP GPU.

The Serpent Rider wrote on 2021-02-11, 00:14:

Bank interleaving is not dual-channel...

Oh, the ServerWorks HE-SL can only do bank interleaving? I thought it was dual channel. Hasn't memory bank interleaving been around since the 486, or even 386, days? Why did it take a server grade chipset to bring it back?

Standard Def Steve wrote on 2021-02-10, 19:25:

Apollo Pro 266T, on the stupid-fast QDI Advance 12T motherboard. Seriously, of all the PIII boards I've played with, this one is easily the fastest. Clock-for-clock, it's around 8% faster than a Gigabyte board based on the same chipset, and it smokes my 440BX, 815, and 694T boards as well.

My retro interests end around Slot 1/2 hardware, which imposes the question - are there Slot1 boards based on A266 ?
I haven't seen such thing. Wonder if there are, but obscured by model/brand name ?

We can run dual Slot boards at high frequencies.
The video above is with P2B-D running at 1575/150 for example.
It can do more, but is limited by clock gen maxing at 150mhz.
But yes, pushing dual CPU systems beyond spec is more effort.

@feipoa
Do you have any metrics of your serverworks board(s) ?
I was always curious about these, but never had the time to examine one.

pshipkov wrote on 2021-02-11, 03:35:

@feipoa
Do you have any metrics of your serverworks board(s) ?
I was always curious about these, but never had the time to examine one.

I don't really have an apples-to-apples comparison of my dual Tualatin setups. It is on my list of things to do, if ever. All I can offer is some numbers on differing hardware that I located in my notes.

3DMark2001SE at 1280x1024x32 in W2K SP4
Intel Intel SAI2 - ServerWorks ServerSet III LE - PC133 - Nvidia FX600 (PCI) - 1.4 GHz - 4345
MSI MS-9105 - VIA 266T - DDR - Nvidia 4400Ti (AGP 4x) - 1.5 GHz - 7126
Tyan S2567 - ServerWorks ServerSet HE-SL - "dual channel" PC133 - Nvidia FX3000 (AGP 1x) - 1.4 GHz - 8675
SuperMicro P3TDDE - ServerWorks ServerSet HE-SL -"dual channel" PC133 - Nvidia FX3000 (AGP 2x) - 1.4 GHz - 9092

3DMark2001SE at 1280x1024x32 in XP SP3
ASRock 939Dual-SATA2 - ULi M1695 - dual channel DDR333 - ATI HD4350 (AGP 4x or 8x) - Opteron 185 at 2.6 GHz - 15666
ASRock 939Dual-SATA2 - ULi M1695 - dual channel DDR333 - ATI HD4350 (AGP 4x or 8x) - AM2-6400+ at 3.2 GHz - 16620
Intel S875WP1-E - Intel 875P chipset - DDR400 - Nvidia 7900GS (AGP 4x or 8x) - Prescott 3.4 GHz - 20702

General observations:
a) PCI graphics is lousy compared to AGP.
b) Is the FX3000 that much faster than a 4400Ti, or is the "dual channel" SDRAM coming into play here?
c) Is the 2x AGP on the SuperMicro responsible for the increase in performance compared to the Tyan, or just some BIOS optimisation?
d) The P4 Prescott smashes the AM2 sytem - mostly due to graphics card difference?
e) Why didn't the Opteron 185 (FX-60) perform substantially better than the Tualatin? Shouldn't it be at least 2-3x faster?

As noted, this isn't really a proper comparison by any means and is not intended to be. It was never my intention to published mix/matched hardware results like this, but this is all I have at the moment.

So the 840 in stock speed tests wins but the BX Wins overall if you overclock. So many of the scores are considered toss ups, at least by "gamers nexus standards" That there is no clear winner in a lot of tests however there are clear losers like i820 and i815 are terrible.

I would of liked to see some other games besides Quake bench marked so it seems it favors the apollo 266 in overclocking test. But In stock speeds its more of a toss up in reality. To the point that the conclusion might be if you are not overclocking than it really doesn't matter and if you are overclocking there's enough variance that I'd probably go with the Apollo 266 over the BX if I base my opinion off quake scores alone, which is why id would of liked to see maybe unreal, and some other cpu bench marks like prime95, maybe throw in an ecoding benchmark like DIVX, 7zip or mp3 encoding tasks. More real world type tests and less synthetic, great read and thanks.

feipoa wrote on 2021-02-11, 00:03:

When I tested this some years ago, the dual channel SDR did indeed show better memory benchmark results compared to DDR w.r.t. Tualatins at 1.4 GHz.

probably has something to do with latency. SDR can run at CL2 and DDR at CL3. If you had an amazing stick of DDR it might not matter much, but for general purpose its just running at the same speed with more latency.

Although the charts all say its running a cas 2 and 2.5 i kinda suspect that the Tras values are different. Wouldn't mind a clarification on that. Unless you run the sdram at 2-2-2-5 Since those values actually matter more than the CAS timing does. I mean if your ram is CL2 but your Tras is like 6. Than your not running the fastest timings. your running at like 2-3-3-6 or even worst 2-3-3-7 which does matter. Because that's basically DDR latency at that point which runs at 2.5-3-3-6 and any advantage you have over lower latency is gone.

when ram runs at the same speed the lower cas time should win, but giving the variance in the tests its hard to tell if the boards were using the tightest timings.

Oh, the ServerWorks HE-SL can only do bank interleaving? I thought it was dual channel.

On a second look, it seems that they've really made a 128-bit SDRAM memory controller, which is a mystery why they even bothered.

Then again, on some motherboards they claim "Four-way memory bank interleaving" and memory can be used only in groups of four - https://www.supermicro.com/products/motherboa … P3/HE/S2QE6.cfm

Warlord wrote on 2021-02-11, 06:05:

I would of liked to see some other games besides Quake bench marked so it seems it favors the apollo 266 in overclocking test. But In stock speeds its more of a toss up in reality. To the point that the conclusion might be if you are not overclocking than it really doesn't matter and if you are overclocking there's enough variance that I'd probably go with the Apollo 266 over the BX if I base my opinion off quake scores alone, which is why id would of liked to see maybe unreal, and some other cpu bench marks like prime95, maybe throw in an ecoding benchmark like DIVX, 7zip or mp3 encoding tasks. More real world type tests and less synthetic, great read and thanks.

Here are more benchmarks from my Apollo 266T. Sorry, they're all overclocked results (1628/FSB-155). I didn't wait 4 freakin' years for another QDI Advance 12T to show up on ebay (and pay through the nose for it), only to run it at a puny 1400/133. 😜
These were all done under XP Pro or MCE2005, depending on when I ran the benchmark, using a 6800GT overclocked to 385/540):

SuperPi 1M: 1m 22.479s
7-Zip 9.2 Total Rating: 1318 MIPS
AIDA64 Memory Read: 1252 MB/s Write: 1186 MB/s Copy: 1118 MB/s Latency: 85.6 ns
UT2003 botmatch 10x7 high: 56.4 fps
Doom 3 timedemo01 10x7 high: 50.8 fps
Expendable 640x480x32: 150.71 fps
Crysis timedemo 800x600 Low: 21.325 fps
3DMark2005 - default: 4597
3DMark2003 - default: 10,070
3DMark2001 - default: 14,059
3DMark2000 - default: 12,132
3DMark99 - default: 12,960 (CPU: 25,021)
AquaMark3 - default: 48,585

Thanks for the info @feipoa.

Good numbers Steve 😉

Apparently i forgot to turn-on SMP in my previous video - meh.

Here is cool Quake3 roll at 258 fps with the same system and SMP:
https://www.petershipkov.com/temp/retro_pc_im … est_run_150.mp4