I'm following this thread.. I was excited to read the last one.. Waiting to see the results so I can make purchases accordingly.. I need build up my retro collection.. The oldest thing I have is a super unstable Socket 4652 nForce 2 Ultra 400 board from BFG-Tech... I had a Tualatin CPU laying around somewhere but sadly since I've moved a lot over the last 8 years, I can't locate it.. 😒

Thank you for the work you're doing!!

dionb wrote on 2020-04-26, 19:04:

One last request: could you see if the board boots without the 6-pin? I actually have a 200W Compaq 24p PSU from 1999. If it might boot with just the 24p I can use that to test it. If it's alive I'll shell out on one of those 375W and 1.5x ATX height monsters 😉

Only tested Post, as I still have to provision the AP550 for the test.

It does POST when the additional 6 pin is not inserted.

The 440bx boards won't run an agp 8x card they are 3.3v electrically which Radeon 9600 and 9800 don't support as well as all ati Xxxx cards

Original Radeon 9800 is 3.3v compatible. Some Radeon 9600 are also compatible.

Cheers for the information. I tested both a 9700pro and the 9800 AIW (which is a Pro, not an XT as I erroneously thought) on the i815ep platform, and both result in a failed boot.

That's probably ASUS being picky.

Good work! Sounds like an interesting project 😀 I am excited to see the results on i840 and Serverworks Serverset III HE-SL chipsets especially since I did not have the chance to include those chipsets in my tests.

Is your Quadro 900 XGL clocked 300 MHz (GPU) 325 MHz (RAM)? If yes, then you should have more or less the same results compared with my GeForce 4 Ti 4600 😉

When it comes to ECS P6S5AT, it takes time to get the best of this board. Finding the fastest AGP GART driver wasn't easy and more importantly, it is very picky on RAM as you already noticed (especially when overclocked). Some sticks won't even start, some will give you very good results some will be poor. I tried like a dozens of sticks before I found the best one. Still I was not able to get it working stable at 166 MHz FSB with Tualatins (more info here ) I hope you will be more fortunate 😉

bloodem wrote on 2020-04-26, 16:16:

I would suggest driver version 30.82 or 29.42 for the Quadro 4 900. They should be faster (same as with GeForce 4 Ti 4xxx).

The mileage varies. I've tested three driver versions on the ECS, since that one is still on the bench:

1P6S5AT		RAM		Type	Mhz	3D2K	3D01	Scpu	Sfpu	Q3a		WHQL
2P6S5AT		1*512		DDR	1400	10638	9646	441	437	168,6		45.23
3P6S5AT		1*512		DDR	1400	11113	9446	436	436	176,1		29.42
4P6S5AT		1*512		DDR	1400	11142	9422	438	437	175.7		30.82

The newest is clearly better in 3D01, whereas the older are better at Quake and 3D2K.

I'll stick to the 45.23, as top 3Dmarks are not the goal of this batch.

The difference performance-wise between the platforms on equal hardware and drivers is what I aim for. Which is not to say that the best of the bench can be tweaked for ultimate Tualatin performance 😀

VooDooMan wrote on 2020-04-29, 10:21:

Good work! Sounds like an interesting project 😀 I am excited to see the results on i840 and Serverworks Serverset III HE-SL chipsets especially since I did not have the chance to include those chipsets in my tests.

Is your Quadro 900 XGL clocked 300 MHz (GPU) 325 MHz (RAM)? If yes, then you should have more or less the same results compared with my GeForce 4 Ti 4600 😉

When it comes to ECS P6S5AT, it takes time to get the best of this board. Finding the fastest AGP GART driver wasn't easy and more importantly, it is very picky on RAM as you already noticed (especially when overclocked). Some sticks won't even start, some will give you very good results some will be poor. I tried like a dozens of sticks before I found the best one. Still I was not able to get it working stable at 166 MHz FSB with Tualatins (more info here ) I hope you will be more fortunate 😉

Thanks! Time is never on our side, but nor was it on the hardware's side, or it wouldn't be called retro 😀.

GPU-Z does not show the ram or gpu clock of my Quadro 900 XGL but according to the net it should be 300/325 as you said.
EDIT: another installtion detected it as 300/650, so exactly the values you anticipated 😀

I'll first do the other mainboards before I try your mod to get it on 150 and 166 FSB.

Testing the VIA chipsets today. What a difference a board makes!

The DFI performed so poorly, I had it replaced by the Aopen AX34 Pro II. I'll adapt the OP.
I suppose there is a reason why they released new revisions of the CA64-TC. I have also one of the newer revisions, but that one suffers from cap rot, so it's an even worse candidate.

Contrary to the DFI, the AX34 Pro II is a splendid motherboard. The Lin Lin with the 1,4 on the AX34 Pro II performs without a hitch and with good ram (CL2), the results are not too far off from the competition. With some tweaking in the BIOS, someone more knowledgeable than me might squeeze a lot more out of this great board. And with the black, it's quite the looker too.

The poor DFI Rev. A:

Black Beauty AX34 Pro II:

Results are in for the Aopen AX34 Pro II [Windows 2K, Nvidia drivers 45.23, Via 4-in-1 AGP driver 1425].

Remark: 150 FSB was only possible with one ram slot filled.

1) At FSB 133 , max about 110W (MB, GPU, SATA), 2*256 SDR-SDRAM CL3

3Dmark2K: 10255
3Dmark01: 9105
Q3Ademo four.DM68: 165.2
Sisoft99: CPU 381 MB/s - FPU 391 MB/s

2) At FSB 133 , 2*256 SDR-SDRAM CL2

3Dmark2K: 10366
3Dmark01: 9258
Q3Ademo four.DM68: 169.4
Sisoft99: CPU 400 MB/s - FPU 411 MB/s

3) At FSB 150 , 1*256 SDR-SDRAM CL3

3Dmark2K: 11330
3Dmark01: 9789
Q3Ademo four.DM68: 183.8
Sisoft99: CPU 441 MB/s - FPU 450 MB/s

4) At FSB 150, 1*256 SDR-SDRAM CL2

3Dmark2K: 11477
3Dmark01: 9946
Q3Ademo four.DM68: 188.9
Sisoft99: CPU 453 MB/s - FPU 465 MB/s

EDIT:

as a comparison, the appalling results of the DFI CA64-TN Rev. A

3Dmark2K: 8943
3Dmark01: 8119
Q3Ademo four.DM68: 138.5
Sisoft99: CPU 235 MB/s - FPU 226 MB/s

I'll check later if BIOS updates can crank up these results.

Results are in for the Chaintech 6VJD [Windows 2K, Nvidia drivers 45.23, Via 4-in-1 AGP driver 1425].

Very stable and fast board at stock speed. Bit finnicky at OC. When experimenting with different ram, I encountered boot problems into Windows. When the system booted successfully, it usually was 3Dmark01 which could not finish and freezed up the pc. Finally, it tok me 8 different batches of ram, a reinstall of Win2k and countless restarts to get the results at FSB 150, Dram Command Rate 1 and CL2,5. At that speed, even my best DDR ram would result in failure at CL2.

Finally settled with one stick of 512 MB Samsung K4H560838F TCCC based ram.

1) At FSB 133 , max about 110W (MB, GPU, SATA), 1*512 DDR-SDRAM CL2

3Dmark2K: 10701
3Dmark01: 9849
Q3Ademo four.DM68: 170.4
Sisoft99: CPU 344 MB/s - FPU 380 MB/s

2) At FSB 150 , max about 116W (MB, GPU, SATA), 1*512 DDR-SDRAM CL2,5

3Dmark2K: 11971
3Dmark01: 9861
Q3Ademo four.DM68: 194.7
Sisoft99: CPU 416 MB/s - FPU 463 MB/s

Remarkable how on this platform 3Dmark01 did not profit from the higher FSB, but Quake clearly did.

On to the RDRAM-platform!

The results of the i820 systems are in:

1. The Aopen AX6C, an early 3-DIMM i820 solution, 2*256 PC800 RDRam + 1 C-RIMM

3Dmark2K: 8546
3Dmark01: 7406
Q3Ademo four.DM68: 140.6
Sisoft99: CPU 345 MB/s - FPU 378 MB/s

Quite the disappointment.

2. The Iwill Dual Slot 1 is a more mature system. I've tested just one 1,4 on a Slot-T adapter, although the Powerleap also works problem-free. It did not want to boot with 2*256 RDRam. So I did the tests with 1*256 and 1*128 (only two DIMM-slots).

3Dmark2K: 9803
3Dmark01: 8767
Q3Ademo four.DM68: 159.4
Sisoft99: CPU 446 MB/s - FPU 470 MB/s

Better, but still not worth the premium price RDRAM fetched in '99. Even the 1,4 Ghz, the final Tualatin released in 2002, does not benefit much from the larger bandwith RDRAM gives. The Tualatin is far better off with fast latencys.

and the only i840 I have:

3. Compaq workstation AP 550, i840, 4*128 RDRam

3Dmark2K: 10239
3Dmark01: 9540
Q3Ademo four.DM68: 162.8
Sisoft99: CPU 568 MB/s - FPU 589 MB/s

A lot better, and impressive results on the memory front.

Overclocking on these boards was either not possible due to BIOS restrictions (Compaq) or unstable, even at the lowest overclocking settings. 1575 (150 FSB) was impossible to achieve on both my i820 boards.

Before posting my final conclusions, we'll visit each board in detail:

1. Asus P3B-F (Intel i440BX)

Modifications: alternative BIOS, accepting the 1,4S Tualatin by default
BIOS alterations: CAS 2 + FSB 133
RAM used: MTec PC-166 SDR-SDRAM

The good stuff:
- ISA-slot
- Rock solid reliability, even at 133 and 150 FSB
- the i440BX is the universally accepted chipset, so no driver issues

The not-so-good stuff:
- Need to OC the NB for 133Mhz operation
- AGP 2X
- VRMs get very hot when running a Tualatin, I added cooling fins
- ATA-33

2. ECS P6S5AT (SIS 635T)

Modifications: replaced all caps
BIOS alterations: CAS 2,5
RAM used: OCZ 512 DDR PC3200 OCZ4001024PDC-K

The good stuff:
- DDR!
- Good overall performance

The not-so-good stuff:
- Original capacitors will have gone bad by now
- SIS chipset not widely supported
- Difficult to source
- Your ram may not work
- No ISA
- No OC with original BIOS or additional fiddling

3. Compaq AP440 (intel i840)
Modifications: 1,4S on Upgradeware Slot T
BIOS alterations: none
RAM used: 4*128 Samsung RDRAM PC800

The good stuff:
- Stellar memory performance
- Very solid, dual CPU capable

The not-so-good stuff:
- Single Tualatin works, but gives BIOS F1 continue prompt
- dual Tualatin demands tinkering
- difficult to source
- needs compaq PSU
- no OC

4. Aopen AX6C (intel i820)
Modifications: 1,4S on Upgradeware Slot T
BIOS alterations: none
RAM used: 2*256 Samsung PC800 RDRAM

The good stuff:
- it's green

The not-so-good stuff:
- probably an early i820, three RDRAM slots
- slow
- OC virtually impossible

5. Asus TUSL2-C (Intel i815)
Modifications: none
BIOS alterations: modded BIOS for AGP4X - CAS2
RAM used: 2*256 PC166

The good stuff:
- Fast
- Great OC performance
- Fits large socket 462 coolers

The not-so-good stuff:
- max 512MB SDR-SDRAM

6. DFI CA64TN Rev.A (Via 694T)
Modifications: none
BIOS alterations: CAS 2
RAM used: 2*256 PC166

The good stuff:
- practical front panel connectors
- max 1,5GB ram
- ISA slot

The not-so-good stuff:
- slowest of them all
- laggy

7. Aopen AX34ProII (Via 694X)
Modifications: 1,4S on LinLin
BIOS alterations: CAS2
RAM used: 2*256 PC166

The good stuff:
- Very fast for a VIA
- Great and stable OC performance
- It's black!

The not-so-good stuff:
- no ISA

8. Iwill DS-133R (Intel i820)
Modifications: 1,4S on Upgradeware Slot-T
BIOS alterations: none
RAM used: 1*256 PC800 + 1*128 PC800 Samsung RDRAM

The good stuff:
- Mature i820
- Dual capable, but not with Tualatins

The not-so-good stuff:
- Strangely no post with 2*256 RDRAM
- Poor stability at OC

9. Supermicro P3TDE6 (Serverset III HESL)
Modifications: none
BIOS alterations: CAS2
RAM used: 2*256 Buffered ECC PC150

The good stuff:
- Impressive memory performance
- Dual Tualatin out-of-the-box
- SCSI U160 onboard
- PCI-X, because longer is better

The not-so-good stuff:
- AGP only stable at 1X, 2X causes 3DMark01 to crash
- No OC whatsoever

10. Chaintech 6VJD (Via Apollo Pro 266)
Modifications: 1,4S on LinLin adapter
BIOS alterations: CAS2,5
RAM used: 1*512 PC3200 OCZ4001024PDC-K

The good stuff:
- Very fast
- Stable OC, fastest result in Q3A on FSB150

The not-so-good stuff:
- difficult to source
- no Tualatin support without adapter

The results, in chart form:

Common equipment: Promise TX4Plus SATA300 + OCZ64GB SSD; Quadro 900 XGL (300/325)
CPU: various Tualatin 1400/512/133
OS: Win2K, Nvidia 45.23

Memory performance: The super expensive i840 is clearly the fastest. Dual SDR-SDRAM takes the second spot and overtakes the theoretically equal DDR-SDRAM solution.

In Quake, the DDR-solutions win, with the Aopen AX34 showing impressive results for what was then called the "cheap" solution.

Quite clearly, the DDR-boards rule on 3DMArk performance. The Apollo Pro is not hindered in these tests by its relatively poor Sisoft performance.

And overall on stock speed:

First prize for the i840 on overall performance.

Silver and bronze for the SIS635T and the venerable i440BX.

EDIT: changed the column labels to Mainboard type instead of the chipset, as requested 😀

Last post with results:

At 150 FSB, only 4 boards were able to complete the test:

Asus P3BF
Asus TUSL2-C
Aopen AX34ProII
Chaintech 6JVD

In Q3A, the Apollo Pro takes the lead, just barely missing 200 fps.

In Sisoft Sandra 99, the BX at 150 FSB is the winner

In 3DMark, the BX again wins, not hampered at all by its AGP2X

Overall, the "old" BX is, at 1575MHZ / 150FSB, the best chipset I've tested. I'd recommend good cooling on the NB, the VRMs and the CPU if ever you plan on having this as a gaming rig.

I think there's a typo in the first charts, you got i820 twice and no i815 ^^

great job though!

edit: ah, there is quite a bit of difference between the two 820s though, would help to directly see at first glance which is which.

imi wrote on 2020-05-08, 12:36:

I think there's a typo in the first charts, you got i820 twice and no i815 ^^

great job though!

I've tested 2 i820, one early implementation (AOPEN AX6C), one more recent (Iwill DS133-R). The slowest is always the Aopen.

The i815 is covered by the i815ep (TUSL2-C). No difference between the i815 or the ep (apart from SB and the absence of VGA).

edit: thx 😀

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...