Having faster memory chips does not guarantee higher clocks. Termination and layout play a large role as well. If the layout sucks or the DDR signals are not terminated properly, you'll have a hard time going faster than the current clock. Team red cards usually need parallel termination on the signals for anything significantly higher than ~220MHz (your 9600TX for example has parallel termination with VTT in place, the Radeon 9000 non-pro does not, the pro does.

I'd even question if the PCI variant of this card would significantly benefit from the additional clocks. Doubling the bandwidth by going to 128bit is more substantial here.

I'd just go for same speed memory and fully populate it.

If you have the ability to do it I don’t see the harm in giving it a go. To me that is your biggest issue.

For any ram upgrades just be sure the pinouts, voltages and ram configuration is the same

Having that extra bandwidth does make a big difference, I bought a very similar agp fx5200 back in 2003 and it was so slow I ended up selling it and going back to my mx440 until I could afford something faster. The stock FX5200 is so slow, unless you plan on using it in something really not suited for it, you should see an uplift in performance. If you are going to run it in a socket 7 you will really need a k6-2+/3 or 3+ as there is more driver overhead in the fx series vs earlier drivers for say the 2/4 mx pci series or TNT, voodoo 3 etc

FullYes wrote on 2026-10-07, 06:00:

If you have the ability to do it I don’t see the harm in giving it a go. To me that is your biggest issue.

For any ram upgrades just be sure the pinouts, voltages and ram configuration is the same

Having that extra bandwidth does make a big difference, I bought a very similar agp fx5200 back in 2003 and it was so slow I ended up selling it and going back to my mx440 until I could afford something faster. The stock FX5200 is so slow, unless you plan on using it in something really not suited for it, you should see an uplift in performance. If you are going to run it in a socket 7 you will really need a k6-2+/3 or 3+ as there is more driver overhead in the fx series vs earlier drivers for say the 2/4 mx pci series or TNT, voodoo 3 etc

Good point. Whatever memory you use: VDDQ and VDD must match your cards current voltage OR you will have to adjust it. By the looks of the card there is only one linear regulator for the memory. The switching one is for the GPU. The Winbond Chips you currently use have VDDQ = VDD = 2.5V.

So you either have to find memory with the exact same arrangement (VDDQ = VDD = 2.5V) or you'll have to adjust the voltage regulation for memory to increase / decrease the voltage. But you'll be probably stuck with VDDQ = VDD since the PCB does not allow for split rails here. The memory used on the 9600TX from Medion/MSI is Samsung K4D261638E-TC36 - it'll run at VDDQ = VDD = 2.5V.

NorsteinB wrote on 2026-10-07, 05:02:

so I would "only" need to populate the vacant RAM solder masks and their surrounding resistor pads.

Let’s be honest, it’s a bit of a tedious task.
Because having to identify and position all those resistors and capacitors around the memory – it’s a bit of a faff.
If you fancy it, why not?

NorsteinB wrote on 2026-10-07, 05:02:

Hi! […]
Show full quote

Hi!

I was a bit disappointed with my Geforce FX 5200 PCI (128 MB, 64 bit) that is still the second fastest PCI graphics cards I own, next to a CM3 ATI Radeon 9000 256 MB 128 bit that sadly lacks support for 8bit paletted textures, table fog and DX9, with the former two being legacy features often used in W9x games and the latter necessary for some Glide wrappers, so if the FX 5200 was 128 bit, it would tick all the boxes (together with EMBM that the Radeon also supports) and be as fast or even a bit faster than the Radeon.

Then I saw Bits und Bolts upgrading the memory of a Geforce 4 MX 440 SE from 64 bit to 128 bit. He even had to jump through some extra hoops because the only memory he had on hand had twice the capacity so he had to make them be recognized as only half that. No idea why someone would do that to an AGP stinker when the 128 bit version is readily available for around 15 bucks but still.

https://youtu.be/ywnIQXl-Mvs

And so I checked: my Zotac Nvidia Geforce FX 5200 PCI 128 MB 64 bit uses the exact same PCB as the 256 MB 128 bit version, and all the resistors around the back of the GPU look the same. It auto detects the memory so I would "only" need to populate the vacant RAM solder masks and their surrounding resistor pads.

The attachment Zotac Geforce FX 5200 128 MB PCI.jpeg is no longer available

And what's more, I kinda pulled the jackpot with a very late chip from 2008, this is a die shrunk GPU and a later stepping, practically guaranteed to run 300-325 MHz instead of 250!

The attachment Zotac Geforce FX 5200 128 MB PCI front.jpg is no longer available

That got me thinking: why bother with 5ns memory, then? Shouldn't the card handle 4 or even 3ns memory instead, as long as the memory chips that are already on there got replaced, too? People do it for 3dfx cards all the time. And with this FX 5200 having the potential to become more or less the ultimate Windows 9x PCI graphics card, it would definitely be worth the hassle. I even have a donor card here, the Medion ATI Radeon 9600 TX (a 9500 Pro with TSOP memory and slightly higher base clock) with 3.6 ns memory!
I believe at one point I even had a Radeon 9600 XT 256 MB with TSOP 3.3ns memory but I gave that one to a friend in Italy.

So anyway, where is the error in my thought process?

And if there is none, should I keep searching for a 256 MB TSOP2 RAM donor card first?

Thats a cheeky hardware hack and I like it.
Ive often thought about doing similar with motherboards that have less features for what looks like no other reason than they didnt fit the controller chip and pin header (firewire ports etc)

Watching this thread.

tehsiggi wrote on 2026-10-07, 05:51:

Team red cards usually need parallel termination on the signals for anything significantly higher than ~220MHz

The FX5200 presents quite an interesting situation.
As with the XFX, this board features series terminating resistors for the memory signal lines.
https://theretroweb.com/expansioncards/s/xfx-pv-t34k-uak5
And on this board, they’ve run out of those resistors in stock.
https://theretroweb.com/expansioncards/s/xfx- … -t34k-nahg-v3-3

Albatron decided not to settle for half-measures and solved this problem in a radical (and cheap) way.
https://theretroweb.com/expansioncards/s/alba … eforce-fx5200ep
https://theretroweb.com/expansioncard/image/i … 72173776690.JPG

It seems that this Zotac graphics card we’re talking about won’t set any memory speed records, regardless of which memory chips are used.
But if it’s a 64-bit card, it’s worth adding more memory chips to increase the memory bandwidth.

Found a better donor - an ATI Radeon 9800 SE 256 MB in 128 bit design.

It cannot be unlocked to the full card anyway, and is basically a worse 9600 XT.

But it has 8 Hynix HY5DU561622CT-4, the correct organisation and voltages.

The maximum clockspeed is slightly lower but if I'm not going to reach far anyway, having twice the memory size might be good for something.

I don't see a scenario for that, but at any rate I would miss the 9600TX more. It is the faster card, even. And less power hungry to boot.

NorsteinB wrote on 2026-10-07, 15:14:

Found a better donor - an ATI Radeon 9800 SE 256 MB in 128 bit design. […]
Show full quote

Found a better donor - an ATI Radeon 9800 SE 256 MB in 128 bit design.

It cannot be unlocked to the full card anyway, and is basically a worse 9600 XT.

But it has 8 Hynix HY5DU561622CT-4, the correct organisation and voltages.

The maximum clockspeed is slightly lower but if I'm not going to reach far anyway, having twice the memory size might be good for something.

I don't see a scenario for that, but at any rate I would miss the 9600TX more. It is the faster card, even. And less power hungry to boot.

You can't unlock the full 8 pipes on the 9800SE? Palit had the same card with all 8 PP enabled as Radeon 9800Pro 128Bit. If you need bios to try it, I have that card somewhere.

With 8PP, i'd take it over the 9600TX probably.. - but it's your call, both have memory that should be suitable.

There are plenty of dead cards out there with the chips you need surely?

Ebay is usually awash with such cards being sold as "untested", "spare/repair", "dont know if it works" etc.

At the very least I didn't find one on localads or german ebay immediately.

I had some Connect3D or whatever Radeon 9600 XT with beautiful TSOp2 packaged 256 MB of 3.3ns memory, but I gifted it to a guy from Naples last summer, who had never had a 9000 series Radeon card in his life!

I don't regret giving him the card but in hindsight I could have just as well given him a BGA equipped model. Oh well, no way to foresee that.

NorsteinB wrote on 2026-10-08, 05:08:

At the very least I didn't find one on localads or german ebay immediately.

I had some Connect3D or whatever Radeon 9600 XT with beautiful TSOp2 packaged 256 MB of 3.3ns memory, but I gifted it to a guy from Naples last summer, who had never had a 9000 series Radeon card in his life!

I don't regret giving him the card but in hindsight I could have just as well given him a BGA equipped model. Oh well, no way to foresee that.

I have a dead 9800Pro 128Bit from Palit, it's dead for what I assume GPU reasons. If you want, I can strip the memory off for you and ship it to you, as it sounds you live in Germany.

tehsiggi wrote on 2026-10-08, 06:53:

NorsteinB wrote on 2026-10-08, 05:08:

At the very least I didn't find one on localads or german ebay immediately.

I had some Connect3D or whatever Radeon 9600 XT with beautiful TSOp2 packaged 256 MB of 3.3ns memory, but I gifted it to a guy from Naples last summer, who had never had a 9000 series Radeon card in his life!

I don't regret giving him the card but in hindsight I could have just as well given him a BGA equipped model. Oh well, no way to foresee that.

I have a dead 9800Pro 128Bit from Palit, it's dead for what I assume GPU reasons. If you want, I can strip the memory off for you and ship it to you, as it sounds you live in Germany.

Great! That way no living 9800 Pro 128 bit must be sacrificed!

Since you've been hanging on to your 9800 card, I can give you a basically identical but working card for your trouble.

It's made by Innovision, which is the same company as Palit, it should be identical save for the sticker on the fan.

I've shot you a DM!

Hi there, any updates on this project?
I basically have the exact same FX5200 PCI from Zotac, but alredy in a 128bit configuration. Probably also a later one, as it came in a box with "Windows Vista ready" written all over it.
I would be interested in seeing how much can you push the clocks and timings on this card before it starts artifacting. Mine came with stock 333mhz memory, but I was able to bring it all the way up to 400mhz without major issues (some minor random artifacts in 2D sprites), finally settling to 375mhz. Didn't touch GPU clocks, but those would also be interesting in see how far they can be pushed.

Mondodimotori wrote on Today, 10:10:

Hi there, any updates on this project?
I basically have the exact same FX5200 PCI from Zotac, but alredy in a 128bit configuration. Probably also a later one, as it came in a box with "Windows Vista ready" written all over it.
I would be interested in seeing how much can you push the clocks and timings on this card before it starts artifacting. Mine came with stock 333mhz memory, but I was able to bring it all the way up to 400mhz without major issues (some minor random artifacts in 2D sprites), finally settling to 375mhz. Didn't touch GPU clocks, but those would also be interesting in see how far they can be pushed.

I got sick and so I didn't yet send the memory and parts to him. I'll do that soonish and then we'll know more once he finds some time to try it out.

Mondodimotori wrote on Today, 10:10:

Hi there, any updates on this project?
I basically have the exact same FX5200 PCI from Zotac, but alredy in a 128bit configuration. Probably also a later one, as it came in a box with "Windows Vista ready" written all over it.
I would be interested in seeing how much can you push the clocks and timings on this card before it starts artifacting. Mine came with stock 333mhz memory, but I was able to bring it all the way up to 400mhz without major issues (some minor random artifacts in 2D sprites), finally settling to 375mhz. Didn't touch GPU clocks, but those would also be interesting in see how far they can be pushed.

tehsiggi is currently sick and I'll be gone for a week starting tomorrow.

But please tell me more about your card!

I take it has 256 MB total capacity then?

The 4 ns memory that is about to go on mine (which is already 50 MHz / 100 MHz DDR faster than the 5ns memory originally on the card) will most probably not even reach 300 MHz.

EDIT: I just randomly remembered I have a defective Radeon 9600 XT with 3.3ns chips. It is only 128 MB but the ICs should be compatible.

As explained above, I sent my "good" 256 MB TSOP 9600 XT to a guy in Naples. IIRC I had tested the 128 MB TSOP 9600 XT but it wouldn't output a picture at all, so most likely the RAM is fine but something else isn't and I couldn't be bothered to search for the cause since I have this extremely cool Powercolor ATI Radeon 9550 with 2.8ns BGA RAM and a Molex 4pin connector that is not needed at all, and it is a BEAST for overclocking, I didn't bother to verify but it trades blows right up with the top 20 scores on HWBOT for the RV350. Don't remember the exact clocks and I'm not much of a tweaker, but it was really surprising me. Another surprise was that the Molex connector did not seem to do a thing at all. Was not needed to run the card, did not improve OC headroom.

Anyway, since this card is as dead as the 9800 Pro 128 bit, I might as well take that RAM - sure it has a lower total capacity, but both the old Winbond W8425GD5H-5 and the new Samsung K4D261638F-TC33 are 16 bit per IC and VDIMM = VDDQ = 2.5V (oddly enough the faster Samsungs appear to have a higher tolerance of .25 vs .1 V). So while my total RAM capacity would stay the same, the rated memory speed would increase dramatically, and if your card is able to actually clock that high, then I can at least hope to run my card at 300/300 MHz.

Which would still be polishing a turd, but hey - that is all I can do for my fastest ISA slot system.

NorsteinB wrote on Today, 14:00:

tehsiggi is currently sick and I'll be gone for a week starting tomorrow. […]
Show full quote

tehsiggi is currently sick and I'll be gone for a week starting tomorrow.

But please tell me more about your card!

I take it has 256 MB total capacity then?

The 4 ns memory (which is already 50 MHz / 100 MHz DDR faster than the 5ns memory originally on the card) will most probably not even reach 300 MHz.

My only alternative right here is 3.6 ns., for that I'd have to kill an OEM Medion Radeon 9600 TX (9500 Pro with TSOP RAM, timings are more casual but clocks are a bit higher in comparison).

I looked everywhere for a Radeon 9600 XT TSOP but could only find 5ns TSOP or 3.3ns BGA. So the only one I had went to a mutual in Naples I guess.

Of course! I got it by random chance on a local ebay listing (Italy), sold as basically brand new in a box, maybe it was late spring or this early summer? 256mb clocked at 333mhz (instead of the 400mhz of the reference model), but the 128 bit bus meant a nice jump in performances from 64bit Quadro 280 and Radeon 9250, even with slower clocked memory.
I tested some overclock on the memory after reading forum posts of the period, with people getting nice OC on FX5200s, both on memory and core.
Of course, since it's a passive card, I didn't want to push it too hard, also the system I use it in is a 462 socket that's compatible only with Thunderbird CPUs at most, currently running a 1200C (instead of the extremely warm 1400C).
I stopped at 400mhz on the memory, with around 5700 points in 3DMark 2000 from 52oo points with stock memory. But I've noticed that after reaching 375mhz you hit a pretty hard ceeling, with diminishing returns, probably hitting a bottleneck somewhere else. Thus, to avoid any kind of artifacts (got some weird pixels in Max Payne loading screen) I stopped at 375mhz.
I could try pushing the core 5mhz at a time but, as said before, passive card in small case.

I'm interested in this more to see if I can convert some of those 64bit PCI cards I have to 128bit ones, since they are extremely rare and expensive nowadays.

Edited the post before

NorsteinB wrote on Today, 14:34:

Edited the post before

Yup, read the update. If you want, I can try to provide you pictures of my memory chips. It sure feels like "polishing a turd" with these cards but... For PCI only systems there isn't really any other option.

A plus is the support for driver 45.23 on Windows 9x (I'm running it on WinME currently), meaning pretty good compatibility for those games.

Oh by the way, could you maybe dump the BIOS of your card?

That and the exact model of your GPU's RAM chips would be helpful for modifying my own one.