Thanks for the extensive documentation of your experimentation!

OVERK|LL wrote on 2025-02-08, 15:45:

Thanks for the extensive documentation of your experimentation!

You are welcome.

One more success: the NSC GX1-266 was able to do 333 MHz!

Compared to the GXm from above (the left one in the pictures), this chip is made in 0,18 µm (the GXm: in 0,35 µm), and it needs 1,8 Vcore only (the GXm: 2,9 Vcore).

This GX1 is able to work with a RAM-frequency of 111 MHz/CL2 at CPU-Ratio of 3,0 (133 MHz/CL3 at CPU-ratio of 2,5 is possible, too, sadly Quake II crashes in Windows, so this is unstable).

The benchmark of CPU-Z still shows a FPU-disadvantage compared to the GX1@300 MHz in its database (no idea why).

Finally, a GeForce FX-5200 was able to work on this board, so now there is a DVI-output, and the nature-test in 3DMark 01SE does work now (the other problems in 3DMark o1SE still remains, so there is no benchmark-result). More interesting, in GLQuake there is a big boost (ca. 47 FPS), compared to the GF 2 MX-400 and the GF 4 MX 440SE (ca. 32 FPS).

From my point of view, there is nothing more to be done faster on this system, as the maximum multiplier for the frequency of the MediaGX-CPU is 10,0 (see the attached screen-shot from its datasheet).

next pics

and the last pics

I think I misunderstood something.
Max multi is 10.
Speed you show is 333.
Is it 33fsb?!

Yeah FSB is low.... from what has been said the last couple of pages, I started dreaming of a MediaGX/Geode daughterboard upgrade for a 386DX ISA board, you design it to have the RAM on, and overlay half the ISA slots with half height PCI, plugging into the 386 socket for ISA and on the board cache, which would probably want to be 15ns or top quality 20. .... miiiight work for a 486 ISA board but there just seems to be more "in the way" vs a less sophisticated board.... probably a lot of messing around with BIOS would be required anyway. ... not really a practical idea as there's not really a mountain of identical or similar 386DX boards hanging around gagging for CPU or much point in making them over to fast 486 low pent performance, when you can just go to low end pent much cheaper.

Hmm.
What clockgen is it?
Look for a chip ~1.5cm wide, 2.5cm long, 2x-48 pins possibly with IcS, realtek or Cyprus on it.

myne wrote on 2025-02-09, 03:32:

I think I misunderstood something. Max multi is 10. Speed you show is 333. Is it 33fsb?! […]
Show full quote

I think I misunderstood something.
Max multi is 10.
Speed you show is 333.
Is it 33fsb?!

Yes, the FSB is 33 MHZ only.
Attached they are some excerpts of the GX1-databook from the AMD-homepage (as AMD is the owner of the Geode-brandmark and the Geode-technics since many years).

About the SDRAM-frequency: the GX1-chip only "guarantees" 79 MHz! So if some system can run more like this - yeah, "dont't worry - be happy!"

And, this GX1-chip does NOT support any kind of L2-cache in generally: "Graphics and system memory accesses are supported by a tightly coupled SDRAM controller which eliminates the need for an external L2 cache" - well, "THANKS A LOT" about this!

myne wrote on 2025-02-09, 04:09:

Hmm.
What clockgen is it?
Look for a chip ~1.5cm wide, 2.5cm long, 2x-48 pins possibly with IcS, realtek or Cyprus on it.

The clockgenerator is a PhaseLink PLL52C59-14TSC, as shown in one picture of the previous page.
It can do max. 37,5 MHz - see the attached datasheet.

Next to the clockgen is the blue switch - in the printed table next to it, switches 1 and 2 are always in the same positions. I checked this in accordance to the PLL-datasheet -> yes, both are for the FSB-settings.
The only problem is - the Turbo-mode can not be enabled (so only possible FSBs are 25/27,5/30/33,3 MHz). I will check the frontpanel-connectors, maybe there is a Turbo-option there.

Best guess : pclk is unused.
The chip was chosen because it has enough pci clocks to run the slots and CPU.
A mod might be possible to connect pclk to the CPU enabling 66mhz etc.

If the above is right, one bclk will go to the CPU. Pclk will be grounded.

Fiddly microsoldering would be required to attempt moving tiny resistors/caps around.

myne wrote on 2025-02-09, 08:53:

Best guess : pclk is unused. The chip was chosen because it has enough pci clocks to run the slots and CPU. A mod might be pos […]
Show full quote

Best guess : pclk is unused.
The chip was chosen because it has enough pci clocks to run the slots and CPU.
A mod might be possible to connect pclk to the CPU enabling 66mhz etc.

If the above is right, one bclk will go to the CPU. Pclk will be grounded.

Fiddly microsoldering would be required to attempt moving tiny resistors/caps around.

As this is the only MediaGX-board I have, I would not do any bigger soldering on it.

Sadly, the fronpanel-connector does not offer a Turbo-button/switch (there is only a Turbo-LED - but, what's the sence of it, if there is no Turbo possible?)
In the PLL-datasheet, pin nr. 5 is the turbo-pin. How can it be used to enable the turbo-function?

The turbo alleges only an extra 2.5%.

My understanding is that that chip does 2 clocks. One for pci, one for cpu. There are a few copies of each.
You say it's a 33fsb.
So I think none of the CPU clocks are used. Instead, one pci clock copy goes to the CPU.

Does that sound about right?

You could see if it's in softfsb. That might be able to enable turbo.

Just tested a Voodoo 3 - 2000 (16 MB, PCI) with this system (clear new installation).
Compared to the performance of the GF-FX5200 tested before: slightly better results in GLQuake and 3DMark99MAX, but much lower FPS in Quake II and 3DMark 2000.
The main problem remains: poor hardware-performance of the GX1-controller for RAM- and PCI-bus.

Trying to install a NIC-adapter into the last free PCI-slot (in total using of 3 PCI-devices), the system became unstable (freezes in WIndows). So I had to use an ISA-soundadapter Creative Soundblaster Vibra, and remove the Soundblaster PCI 128.

myne wrote on 2025-02-09, 10:36:

You could see if it's in softfsb. That might be able to enable turbo.

Sadly, softfsb does not support this PLL52C59-14! I just checked the readme-file of it - it's listed there as unsupported.

Hmm. Well. It looks like pin 5 has a resistor on it. Presumably, based on 1 being on, and it linking to pin1 (vdd) it is a pull up resistor. If it measures 1.5,2.5 or 3.3v it would imply I'm right. If so, simply removing it should enable your extra 2.5%

myne wrote on 2025-02-09, 12:26:

Hmm. Well. It looks like pin 5 has a resistor on it. Presumably, based on 1 being on, and it linking to pin1 (vdd) it is a pull up resistor. If it measures 1.5,2.5 or 3.3v it would imply I'm right. If so, simply removing it should enable your extra 2.5%

OK, I will check this until end of next week, thank you. For now, my weekend is sadly over...

You mentioned that the GX1 had a 17% reduction in FPU performance in CPU-Z compared to the GXm at the same frequency. Were these tested on the same system, or are you using the user table from CPU-Z? Have you verified this occurrence in game-based benchmarks on your system?

I've never seen a GXm board with 3 PCI slots. This is interesting. Are you able to use all 3 PCI slots simultaneously for bandwidth heavy tasks, e.g. video, hard drive, and 100 mbit ethernet? EDIT: looks like you tried this before I posted my comment. Your system was unstable with 3 PCI cards. I understand why most systems only come with 2 PCI slots. On my system, I use PCI1: Voodoo Banshee, PCI2: Promise TX2 ATA100. I leave the CD-ROM connected to the motherboard's IDE for DOS CD-ROM access.

Strange that you don't have jumper control over the CPU voltage like on other GXm boards. My board has jumpers going down to 2.1 V.

If you cannot use the fastest SDRAM frequencies with 256 MB, try reducing your total memory size until you can. I needed to use 64 MB to achieve best BIOS SDRAM timings (SDRAM divisor of 3 and refresh 2T) with a stability I was satisfied with. The databook states, "SDRAM frequencies over 79 MHz are only supported for certain types of closed systems, and strict design rules must be adhered to." Thus, 100 MHz is already pushing the system further than intended.

Does your system BIOS have CPU clock to SDRAM divisor options of 2, 2.5, 3, and 4? Mine only has the 3 and 4 option. Are you able to upload your systems BIOS here?

Have you tried using EDO DIMMs with the /2 and /2.5 options? I recall reading somewhere in the datasheets that EDO is supported.

Could you provide readable photos of your SDRAM sticks? I'm surprised you could achieve stability with 256 MB at /3 and 2T at 300 MHz, and later, 333 MHz.

I am curious why you went for the BGA GX1-266 CPU. There are ceramic GX1-300 2.0V CPUs. There was an endless supply of them on eBay at one point.

How much effort would it entail to bodge wire the PLL to run at 34.2 MHz or 37.5 MHz? I bet the TURBO function runs at 34.2 MHz. Try to toggle it manually. EDIT: Looks like another posted suggested how to do this already.

Are you willing to tabulate the benchmark numbers between the GF2MX, FX5200, and Voodoo3-2000 at 300 and 333 Mhz? Do you have a Voodoo3 3000 or 3500? Or can you overclock your V3-2000 to 3000/3500 speeds for a fairer comparison?

Are you able to use the Cyrix companion IDE drivers for your CD-ROM drive within Windows 98SE while still having pre-loaded the DOS CD-ROM driver? For my system, I found that if I load the DOS CD-ROM driver, then try to load Windows, I will get compatibility mode paging on my CD-ROM drive. It is my primary complaint with this system. I have no other system with this issue. If I run a DOS game which has a CD-ROM disc verification check, I need to edit autoexec & config.sys to re-enable the CD-ROM in DOS, the reboot.

myne wrote on 2025-02-09, 12:26:

Hmm. Well. It looks like pin 5 has a resistor on it. Presumably, based on 1 being on, and it linking to pin1 (vdd) it is a pull up resistor. If it measures 1.5,2.5 or 3.3v it would imply I'm right. If so, simply removing it should enable your extra 2.5%

I'm looking forward to this test with great anticipation!

feipoa wrote on 2025-02-09, 21:36:

On my system, I use PCI1: Voodoo Banshee, PCI2: Promise TX2 ATA100.

Can you explain shortly, which system (mainboard, CPU etc.) do you have?

feipoa wrote on 2025-02-09, 21:36:

You mentioned that the GX1 had a 17% reduction in FPU performance in CPU-Z compared to the GXm at the same frequency. Were these tested on the same system, or are you using the user table from CPU-Z?

Just have a look at my screen-shots: both of my used GXm AND GX1 are benchmarked with this lower FPU-performance, compared to the GX1-300-values from CPU-Z (even my GX1@333 scores a bit lower than the GX1@300 from CPU-Z)

feipoa wrote on 2025-02-09, 21:36:

If you cannot use the fastest SDRAM frequencies with 256 MB, try reducing your total memory size until you can...100 MHz is already pushing the system further than intended

I tried this with one piece 128 MB stick (various brandmarks) in various RAM-slots (as I do not have 64 MB sticks), sadly no success. BTW, I was able to reach 111 MHz (not only 100 MHz) with the GX1@333 (see my screen-shot once again).

feipoa wrote on 2025-02-09, 21:36:

Does your system BIOS have CPU clock to SDRAM divisor options of 2, 2.5, 3, and 4?

Yes, it has the full range of CPU-ratios from 2 to 4

feipoa wrote on 2025-02-09, 21:36:

Are you able to upload your systems BIOS here?

It can be downloaded directly here: https://theretroweb.com/motherboards/s/supert … -st-mgxm3a#bios

feipoa wrote on 2025-02-09, 21:36:

Have you tried using EDO DIMMs....

I do not have any

feipoa wrote on 2025-02-09, 21:36:

I am curious why you went for the BGA GX1-266 CPU. There are ceramic GX1-300 2.0V CPUs.

Well, my GX1-266 needs 1,8 V only (not 2,0 V or more) - so I hoped, this lower voltage can do the OC-ing from 266 to 333 MHz easier - and I was right

feipoa wrote on 2025-02-09, 21:36:

How much effort would it entail to bodge wire the PLL to run at 34.2 MHz or 37.5 MHz

I will try to do this until end of this week

feipoa wrote on 2025-02-09, 21:36:

Are you willing to tabulate the benchmark numbers between the GF2MX, FX5200, and Voodoo3-2000 at 300 and 333 Mhz? Can you overclock your V3-2000 to 3000/3500 speeds for a fairer comparison?

Yes. I will try to overclock the V3-2000 using PowerStrip (even I do not expect a big advantage due to the poor benchmark-results with the Voodoo on this system until now in generally).

feipoa wrote on 2025-02-09, 21:36:

Are you able to use the Cyrix companion IDE drivers for your CD-ROM drive within Windows 98SE having pre-loaded the DOS CD-ROM driver?

I am using Windows ME, without pre-installed DOS. On this system, using DOS (and CD-ROM in DOS) is not interesting for me.

Here are the benchmark-results in WinME and DOS (using a bootable floppy disk of Win98SE) for the GX1-CPU working at 300 and 333 MHz and using a Geforce 2 MX-400 PCI, GeForce FX 5200 PCI, Voodoo 3-2000 PCI and Radeon 9250 PCI.

Sadly, still having DirectX 7.1, with the Voodoo 3 I forgot to do the GLQuake-test. After installing of DirectX 9.0c, this VGA does in GLQuake only 7-8 FPS 🙁 As DirectX can not be rolled back, I would not install WinME once again for this single test (as I do not have a copy of the first installation anymore).

The voodoo3 was such a great all-rounder for someone trying to cover as many eras of PC Gaming back then. What good times.
Makes me happy with my K6 III+ system running at 633Mhz with a Voodoo3 3500 AGP TV card.
I found a nifty adapter cable that allows me to plug in a regular DVI display cable in the back of it so that I don't have to use that huge in-line TV output dongle.

Good work on the Benchmarks!