5200 and 5500 are essentially the same GPU so it really doesn't matter which core they are using if its got a 128bit bus and can handle the slightly higher 5500 clocks.

Yes thats all the 5500 is .. its a guaranteed 128bit memory bus and higher clocks from the 5200, there is no other differences AFAIK.

As for fab quality ..well its from China so chinesium quality is to be expected along witht he huge number of cut backs on components the card doesn't need to output a picture. I also suspect the fans are recycled ones from some place that made too many and didn't need them, naturally they are also chinesium quality.

You do get what you pay for here, sure its new but its cheap and from China, I have one of these too and I used it once and promptly put it back in its box and forgot I owned it, I bought a BFG PCI 5500 a bit later and was far happier with that.

Sometimes recycled stuff manufacturers in China make can be great. Like i got some of those LGA2011 boards and while they use recycled chipsets everything else is new, manufacturing quality is great and whole thing, with its hacked hardware and firmware, works surprisingly well. It feels no worse than a board from any major manufacturer, even better than cheap ones can be nowadays.

However what you describe sounds like someone making them basically by hand or with very poor equipment. And the fact they are trying to relabel the chip only tells that they are specialized in scam, not quality manufacturing.

I also do not understand why this cards need to exist at all and why they are pushed so much by certain youtubers. This are at best a workaround to a problem which has multiple better solutions. Sure if someone wants to make specific suboptimal system work they can be handy, but building a system around one of this like some people do... i just do not get it...

Archer57 wrote on 2025-06-22, 04:14:

I also do not understand why this cards need to exist at all and why they are pushed so much by certain youtubers. This are at best a workaround to a problem which has multiple better solutions. Sure if someone wants to make specific suboptimal system work they can be handy, but building a system around one of this like some people do... i just do not get it...

As Mr. Crabs said "i like money". Retro computing has a hype these years, with YouTube promoting it (which in itself isn't bad imho, i like revisiting the past).

Since the technology itself had become relatively cheap, people like to go "overboard" in performance (putting an FX5200 into ancient stuff) or look for cheap alternatives.

Old gpus are cheap to come by, there are people filling the niche.

I'd rather see a fleshed out non cheaped design, than what we've got. On the other end it can be an easy starter for people who have skills and equipment to improve upon.

Nice review, makechu!

Yeah, these FX 5200/5500 chinesium cards can vary quite a bit in quality... like anything else, I suppose.

The different RAM chips and GPU chip sitting slightly crooked seriously made me laugh a little. 🤣 Sounds like someone is making these in their garage as a side-hustle. 😁

Also FWIW, your new card doesn't even reuse polymer caps like you thought it did - no, it uses the worst possible thing: Sacon FZ regular electrolytic caps without safety vents! These literally blow up like popcorn near EOL... and they reach EOL much faster than even cheap Taiwanese cap brands. Actually, what makes them so terrible is they go pretty much open-circuit when they fail... meaning if the VRM on which they sit does not have any capacitative redundancy (e.g. those missing extra ceramic caps you added) and/or if it can't handle the capacity decrease, you could end up with a dead GPU or RAM IC. So with all of the work you did on that newer card, consider replacing those fake FZ polymer caps too.

Archer57 wrote on 2025-06-22, 04:14:

I also do not understand why this cards need to exist at all and why they are pushed so much by certain youtubers. This are at best a workaround to a problem which has multiple better solutions. Sure if someone wants to make specific suboptimal system work they can be handy, but building a system around one of this like some people do... i just do not get it...

Agreed.

Basically, these only make sense if you somehow ended up with a cheap or nearly free P3/P4 machine with PCI slots only and just want to quickly build a Win98 PC rather than put together a more planned-out system.

I actually had a former coworker who got 10 nearly free Dell Optiplex P3 beige desktops (literally the desktops, not the tower version.) Of course these were meant and served as office machines in their previous life, so no AGP slot and not many other expansion options either. So he put 10 of these FX 5500 PCI cards on them and set them up with both W98 and XP. The goal was an extra 10 machines for a large retro LAN party at his house... and these machines with the FX 5500 PCI worked quite OK for that. Blasted through some Quake, Mechwarrior, and a few other games without issues. Framerates weren't the best in some games, but overall well playable.

So I do think these cards do have a place in the retro PC world. But like you, I just wouldn't go out of my way to get one, unless absolutely needed, and certainly not recommend it as something aspiring to get.

Yeah, the use case for these cards is pretty limited, and their build quality is craptastic.

If the target machine is an LGA775 system with both PCIe and PCI slots, I see no reason to settle for the subpar performance of a PCI FX5500. Even a cheap PCIe Radeon X550 or X600 would leave it in the dust. As for the legacy features, the lack of paletted textures only affects a small number of games, while table fog is fully supported on R300 and R400 under WinXP when using Catalyst 7.11. For those running Win98 on LGA775, it makes sense to dual boot with WinXP anyway, so no issues there.

I am definitely going to change those electrolytic / polymer caps for peace of mind, after some reverse engineering what those regulator circuits on it are good for. Way too much effort for what the card is worth, but I don't really like this card to go to e-waste either.

I can also somewhat see a use case for these cards, if there is no AGP (i.e. integrated AGP graphics having devoured the slot), and no PCIe, and there is still a need for somewhat marginally ok speed at higher resolutions, and / or anti-aliasing. But at the current prices for the quality one gets, even that is highly debatable.
The amount of RAM on this is just a gimmick, as I think nothing that would require that much would run at all acceptably on this card. This card seems to already struggle heavily when it needs to process textures amounting to just 32 - 64 megs on the card.

The inability to use 1920x1200@60Hz over DVI is likely due to the timings used being outside of pixel clock limits for DVI . Using reduced blanking to fit within the 165MHz DVI limit should work, AFAIU, though these cards/chips could have other issues/limitations too. Using custom timings under Windows/Linux is possible. Under DOS, a tweaked monitor EDID or EDID emulator wiould likely be needed.

EDIT: At what speed is the RAM clocked versus a reference design? I recalled this AliExpress AGP cards and was wondering if these FX 5500 were hobbled in a similar way.

https://tomverbeure.github.io/video_timings_calculator

darry wrote on 2025-06-22, 11:08:

EDIT: At what speed is the RAM clocked versus a reference design? I recalled this AliExpress AGP cards and was wondering if these FX 5500 were hobbled in a similar way.

On the blue card, which was the problem child card, the RAM seemed to be running at just 133 base / 266 MHz DDR speed. Which is just sad as the winbond chips, according to their datasheets, should be able to do 400 MHz...

edit;
Could be partially caused by leaving out all those capacitors... I do not remember what the better version of the card was clocked at, against which I compared that. The better card seems to be completely identical, besides some very minor changes like having the fuses and their footprints, having way better quality all around, and having the correct-ish looking set of components. The RAMs look to be new Samsung chips on that one.

makechu wrote on 2025-06-21, 23:06:

In 2022, I bought one of these "FX 5500" PCI cards via ebay, and it has worked pretty good for what it is. So, I bought another one this year, and what a letdown it was. It seems there are some extremely cost-cut versions of these cards floating around with significant quality compromises.

Thank you very much for your extremely thorough and detailed analysis! I have been eyeing an FX6200 card on the Aliexpress which costs about 1/3 from the Ebay card that you are investigating. I decided against it. Not only that it had no fan, but also the fan connector socket was missing, which in my (very superficial) opinion was not best practice. Now thanks to your discoveries I see what lies deeper underneath such "new" offers.

makechu wrote on 2025-06-22, 11:32:

On the blue card, which was the problem child card, the RAM seemed to be running at just 133 base / 266 MHz DDR speed. Which is just sad as the winbond chips, according to their datasheets, should be able to do 400 MHz...

Ouch, that is will certainly hurt the card's performance, even on the PCI bus!

Perhaps this was the only stable speed due to using chips from different manufacturers. Or maybe they are using really shoddy leftover reject memory chips. Or maybe it was due to the lack of the ceramic caps. Who knows, right.
With the ceramic caps added, I do wonder whether you would be able to "OC" it back to normal speeds - i.e. 200 MHz / DDR 400M... or if not, then at least 166.
That said, I've noticed that on a lot of cheaper OEMs' cards, the VRAM is almost never running at its full spec speeds for whatever reason. And not only! For example, I have an ASUS Radeon 9200 SE with 5 ns DDR RAM chips (meaning the card should be able to do 200 MHz / DDR400), yet they are set to run at only 166 MHz from the factory.

Anyways, I know it may sound strange, but I really like geeking out on cheap / mundane hardware like this. A lot of times you end up learning a lot of new things, especially when you decide to just "try something out", since the hardware is not really anything special (if not outright junk, or build from, that is. 😁 )

momaka wrote on 2025-06-23, 06:50:

Ouch, that is will certainly hurt the card's performance, even on the PCI bus! […]
Show full quote

Ouch, that is will certainly hurt the card's performance, even on the PCI bus!

Perhaps this was the only stable speed due to using chips from different manufacturers. Or maybe they are using really shoddy leftover reject memory chips. Or maybe it was due to the lack of the ceramic caps. Who knows, right.
With the ceramic caps added, I do wonder whether you would be able to "OC" it back to normal speeds - i.e. 200 MHz / DDR 400M... or if not, then at least 166.
That said, I've noticed that on a lot of cheaper OEMs' cards, the VRAM is almost never running at its full spec speeds for whatever reason. And not only! For example, I have an ASUS Radeon 9200 SE with 5 ns DDR RAM chips (meaning the card should be able to do 200 MHz / DDR400), yet they are set to run at only 166 MHz from the factory.

Anyways, I know it may sound strange, but I really like geeking out on cheap / mundane hardware like this. A lot of times you end up learning a lot of new things, especially when you decide to just "try something out", since the hardware is not really anything special (if not outright junk, or build from, that is. 😁 )

I did a lot of analysis of existing PCBs in the ATI 9X00 Series recently for another project. What you mention can often be seen. Memory speeds well below specification. There are some things to consider here:

a) Timings
Certain memory speeds are only guaranteed at certain timings by the manufacturer. So even if they are 4ns chips, you might not get them running on 250MHz, if the timings are too aggressive.

b) PCB Layout, Termination and Power delivery

Especially cheaper cards like the 9200 Series have their PCB optimized for cost. Meaning that they'll have to only achieve a certain speed. For 9200SE it is 166MHz memory speed, even though 5ns or 4ns chips are used here and there. The layout of those PCBs is not optimized for high speed.

This can also be seen in the way the connections between GPU and memory are terminated. There are two types of termination for DDR signals: parallel and series. Especially with higher memory speed, the signal integrity will suffer if signals are reflected at the end of the wire (e.g., at the Memory IC). This is where termination comes into play. For low end cards, series termination appears to suffice to get stable results. This is usually up to 200MHz of memory clock. You can find Radeon 9000 non Pro, 9200, 9550 and 9600 with low performance memory (200MHz) and only using series termination. Everything higher - at least in the ATI world - used series plus parallel termination for better signal integrity and thus higher clock speeds.

I've verified this by looking at cards I have had laying around:

Radeon 9000 Pro - 250MHz Memory Speed
Radeon 9600 XT - 300MHz Memory Speed
Radeon 9500 - 270MHz Memory Speed
Radeon 9700 Pro - 310MHz Memory Speed
Radeon 9800 Pro - 340MHz Memory Speed
Radeon 8500 - 275MHz Memory Speed
Radeon 8500LE - 250MHz Memory Speed
Radeon 9550XT - 325MHz Memory Speed
Radeon 9600TX - 270MHz Memory Speed
Radeon 9600Pro - 300MHz Memory Speed

All the above mentioned use parallel termination. Without that, I doubt they would ever reach those clock rates.

I even did an experiment on a Radeon 9200SE from HIS. I swapped the ICs from Hynix 4ns to Samsung 3.6ns (also adjusted the power delivery accordingly). The result yielded basically no improvement in overclocking potential, even though the 3.6ns chips ran fine on other cards at 270MHz. Even changing the timings to the others card values did not allow me to run the chips with their default clock.

Of course, the memory interface in the GPU plays a role too, not gonna lie. But the surroundings are imho way more important in this case.

Also the power delivery is terrible on those low end cards. High speed signals have high current peaks in a short time frame. This requires proper capacitance to be present on the power rails at the right spots in order to fulfill the power delivery requirements of memory.
Seeing that stuff is missing here, it will add to the list of potential causes for bad memory performance.

On the "why do they use 5ns chips if they only run at 166MHz" - BOM optimization. If you run 5 different cards, of which 3 need to run at 200MHz and 2 at 166MHz, you'll probably get a better deal (and process cost) if you just use one type of memory IC and just clock it lower.
Nice if you want to overclock to 200MHz on the low end cards (which on my 9200SEs basically always worked).

Another thing: 200MHz and 166MHz (5ns and 6ns) are commodity speeds from the DDR1 era. I have Radeon 9250 SE cards that use just the same ICs as a regular DDR1 memory stick would use, which is extremely cost effective.

tehsiggi wrote on 2025-06-23, 07:44:

This can also be seen in the way the connections between GPU and memory are terminated. There are two types of termination for DDR signals: parallel and series. Especially with higher memory speed, the signal integrity will suffer if signals are reflected at the end of the wire (e.g., at the Memory IC). This is where termination comes into play. For low end cards, series termination appears to suffice to get stable results. This is usually up to 200MHz of memory clock. You can find Radeon 9000 non Pro, 9200, 9550 and 9600 with low performance memory (200MHz) and only using series termination. Everything higher - at least in the ATI world - used series plus parallel termination for better signal integrity and thus higher clock speeds.

I agree that the termination on the RAM chips on this card should be something of the series-parallel sort, more specifically what should have been used, seems to be called Series Stub Termination logic (SSTL) approach, where the parallel resistors are pulled to a regulated voltage, i.e. termination rail, near the logic transition threshold level for the corresponding data & signal lines. This way, the termination also acts as a weak pull-up when the line is being pulled high, and as a weak pull-down when being pulled low.

There can be some limitations against using this if the memory controller can't handle it. But, in this case I am going to make a guess the reason is cost. I have used the required buffer / regulator chips in the past to create the stable voltage which this approach requires, and the chips for doing it were surprisingly expensive, more than 10 € for each rail, mostly because of the magnitudes of currents involved to keep the voltage stable. Don't ask why I know :p

By quick glance, this card indeed only seems to have series resistors on the memory related lines, so even that is cost-cut / suboptimal in the way it is implemented.

I can't really see from the pictures, but it appears that - at least from I can make out - there is no length matching for the memory signals. Is that assumption correct? (I know sometimes that happens under the memory iCs themselves).

Series only termination will probably suffice for 166 to 200MHz. I even had a Radeon 9000 non-pro from Manli that uses no termination resistors at all. It runs memory at 200MHz. I haven't checked the timings but it baffled me a bit.

tehsiggi wrote on 2025-06-23, 18:46:

I can't really see from the pictures, but it appears that - at least from I can make out - there is no length matching for the memory signals. Is that assumption correct? (I know sometimes that happens under the memory iCs themselves).

There are no "squiggles" / serpentine tracks anywhere, and some RAM related traces have vias and some don't. Physically the traces look longer to the RAM chips near the back of the card, than what they look to the chips on the side, so probably very little to no delay considerations are made.

Ouch.. If that's the case, that adds itself to the potential memory speed issues.

I first thought that the issue with the low RAM frequency would be mostly caused by the mismatching of RAM chips. But I have now checked the clocks from the green card too, and it has the same clocks as the blue card, even though all of the RAM chips on that one are at least all the same.

That makes me think it's deliberate. Safe route. 133MHz will probably always work for them, even with a crappy layout/power/memory.

Because it might be of interest to some, I tested a bit of overclocking on the better / green card, which has the Samsung RAM chips that look new. The code on those is K4H561638F-TCB3, corresponding to 333 MHz DDR / 6 ns clock cycle RAM chips.
However, the system has K6-3 @ 550 MHz, so it is heavily CPU bottlenecked for this card, so these results are mostly meaningless regarding any stability considerations.

I used 3DMark 2000 Pro to run the standard set of benchmarks, and I used 1920x1200 resolution & 32 bit colors. I used RivaTuner 2.24 "MSI Master Overclocking Arena 2009" edition to do the "driver level" overclocking, and the stock nvidia driver version was 45.23. The operating system was WIndows 98 SE with unofficial SP3 Lite.

The benchmark tests all went through without any problems. I also monitored the card with a thermal camera, and at most I saw temps reaching 50 - 60 degrees C anywhere on the card when overclocked. The hottest part was the inductor on the voltage regulator. The GPU heatsink and RAM chips stayed pretty cool, at 40 C or below. It should be noted though, that I have drilled a couple of holes through the heatsink near the GPU to let a bit of air to flow under the heatsink and over the RAM chips. Can't say how much that affects things. Also, there was good airflow over the card from a 120mm case fan.

With the overclock (270 MHz on core, 333 DDR MHz on RAM), I got a score of 2066.
Without overclock (250 core, 266 RAM), I got a score of 1777.

Maybe I will test the crappier card later if that can also do the same in that system. Although it has a mismatching set of RAMs, each seem to have headroom, according to their specifications, to run up to 400 MHz. However, currently I do not want to push it with the junk electrolytic capacitors on it, even if I have added all the missing ceramic capacitors. So I might do that test much later, maybe sometime this year if I remember to get around to it...

best fx5200 5500 pci card is a card you create yourself.

i made my own from PNY pcb, used fx5200 ultra gpu and 3.6ns memory chips, i got a 340Mhz core and 600mhz ram and 1st place on hwbot in 3dmark 99,00,01,03 scores -)

https://hwbot.org/hardware/videocard/geforce_fx_5200_pci