VOGONS


First post, by Masterchief79

User metadata
Rank Newbie
Rank
Newbie

Hey guys,
after successfully transplanting the GPU of my first FX1300 onto a FX5950 Ultra, I wanna do more with these cards.

The attachment IMG_20240811_160126006.jpg is no longer available

I want to turn another FX1300 into a PCX5950 (128MB@350/275 -> 256MB@475/425MHz).
This is mainly because I can and I want to. A NV38 based PCI-E card doesn't really have any real benefits to me, I have plenty of AGP systems around. But I'm a collector and semi-professional technician of retro-hardware and today I decided I would like to have a PCX5950. 😁

Here are the challenges I'm looking at:
- The 2.8ns memory chips will need to be swapped. I have two defective FX5950 Ultras here with 2.0ns Hynix mem so I could just swap those chips over. Problem: I would end up with a 128MB card since the Quadro only has 8 memory chips opposed to the 16 on a 5950U (front and back). I want to double the memory if possible.
- Afaik there are no fitting hynix memory chips with deliver 8x32 = 256MB. There are AFM-25 on a 9800XT but they are specified to 2.5ns (400MHz) and would probably not do 425MHz reliably. Might be worth a shot though. I could do the rest with timings.
- not currently sure if other manufacturers like Samsung or Infineon did GDDR ramchips in this configuration - let me know
- Samsung GC20 or GC16 is 144FBGA but is probably not pin compatible, since it's GDDR3 memory
- Voltmod for the GPU (1.35 -> 1,55V) - should be no big issue.
- possibly Voltmod for the memory - should be no big issue.
- better cooling for the GPU and mem Chips (VF700 is available).

Swapping memory modules from a different brand onto the card usually requires you to modify the memory straps. I've never done that before but I'm told usually they show the configuration in the schematics.

Questions at this stage:
- Which card had 256MB GDDR memory in a 8 chip configuration, ideally with clockspeeds higher than 400MHz (<2.5ns)?
- schematics for some 5900 series graphics card, no matter if AGP or PCI-E, or some old Nvidia schematics in general (anything between NV25 and G71) - found some
- information about memory straps and changing them between memory brands
- general input or questions welcome

This thread already helped: PCIE Graphics Card and it had a schematic inside which I attached below.

New question: Since the PCX5900 is GDDR-based, and the PCX5950 allegedly GDDR3-based:
- is the pinout compatible? There seems to be some differences but there's also a lot of similar stuff
- is the NV38 GPUs memory controller able to handle GDDR3 memory? Wasn't there a FX5700 Ultra GDDR3 at one point? Is the PCB identical? That's comparing to GDDR2 though
Information on that: https://hothardware.com/reviews/nvidia-geforc … -with-gddr3-ram

Currently looks like GDDR2 and GDDR3 are pin-compatible but GDDR is not to either of those two.

So it looks like the only feasible option is to get some Hynix AFM-25 memory and hope it clocks to 425MHz. I have a buddy who has some defective 9800XTs. I don't know where else I could steal it from - input welcome.

Next question: What am I going to do BIOS-wise? I'm not too familiar to modifying BIOS files outside of Nibitor and the like. Techpowerup doesn't host files that old and even if they did, I'm not sure there are original PCX5950 files anywhere (if they even existed). The problem is memory size of course, everything else should be easily modified. I think I'm gonna need help on that too.

Reply 1 of 5, by The Serpent Rider

User metadata
Rank l33t++
Rank
l33t++

PCX5950 never existed.

I must be some kind of standard: the anonymous gangbanger of the 21st century.

Reply 2 of 5, by Masterchief79

User metadata
Rank Newbie
Rank
Newbie

I can still make one as long as someone helps me with the BIOS and I figure out if the 9800XT chips are compatible.

At least a PCX5900+ with 128MB 500/500 should be doable relatively easily.
I got my quadro today, it clocks really well. 450MHz+ on default settings. Memory tops out at about 400MHz but that's not a problem since I'm gonna swap that anyway.

Reply 3 of 5, by Masterchief79

User metadata
Rank Newbie
Rank
Newbie

I swapped over the 2ns chips from a 5950 Ultra for now, did the BIOS and strap mods and the card is recognized and works as a PCX5900 - the closest I will get I think. So no doubling memory capacity. But I want full 5950U clockspeeds on this card at 475/475 instead of 350/275.
Cooling and voltmods are also (partially) done.
The big problem right now is the memory rail on the card. Sometimes it shuts the card off, I think because of OCP. I want to lower the memory voltage but it's not that easy with this ISL6534 controller.

The attachment k-excel.jpg is no longer available

Reply 4 of 5, by Masterchief79

User metadata
Rank Newbie
Rank
Newbie

Update: Got the card finally working at full 475/475 clockspeeds today.
To recap: After swapping Hynix AF2 2.0ns memory from a 5950 Ultra over to this Quadro card, it started crashing under load.

The problem is, the default memory voltage for the Quadro FX1300 sits at 2.85V, which is too high. Even 2.8ns memory is specified for 2.5V, as are all other bins. No idea why they set it so high.

The higher memory VRM load triggers some sort of OCP which results in 0V GPU and memory
It's important to note that the controller ISL6534 does NOT actually have OCP or OVP protections. I determined that due to the higher VRM load with faster memory, the 2.8V sometimes spikes to above 3.2V in the nanosecond-range.
Why is that a problem? Because the controller has a internal referance voltage of 3.3V and all VOUT rails MUST be lower than 3.3V, or the controller detects a sort of "short circuit" scenario (might not be strictly accurate but close enough).

By having a good look at the Gigabyte PCX5900 schematics, which use an identical PCB, I was able to reverse-voltmod the memory down to 2.45V which is much closer to the specification and is also plenty to run at 475MHz. The trick here is that the corresponding feedback divider resistor network isn't connected to the FB2 pin, but to the VREF Pin (10) of the controller.

TLDR: Lower memory voltage (complicated), no more protection triggered, card finally able to clock freely.

3D Marks attached. Might be the first PCI-E NV38 card with full 5950 Ultra clockspeeds, except maybe some PCX5950 sample boards.

The attachment IMG_20241110_221217746.jpg is no longer available
The attachment IMG_20241110_221200428.jpg is no longer available
The attachment IMG_20241110_221144613.jpg is no longer available
The attachment PCX5900 475-475.JPG is no longer available
The attachment PCX5900 475-475 01.JPG is no longer available

ToDos left:
Make Voltmod for GPU permanent
Test card in x16 system with faster CPU and proper drivers and compare to 5950U AGP

Reply 5 of 5, by Dothan Burger

User metadata
Rank Member
Rank
Member
Masterchief79 wrote on 2024-11-10, 21:19:
Update: Got the card finally working at full 475/475 clockspeeds today. To recap: After swapping Hynix AF2 2.0ns memory from a 5 […]
Show full quote

Update: Got the card finally working at full 475/475 clockspeeds today.
To recap: After swapping Hynix AF2 2.0ns memory from a 5950 Ultra over to this Quadro card, it started crashing under load.

The problem is, the default memory voltage for the Quadro FX1300 sits at 2.85V, which is too high. Even 2.8ns memory is specified for 2.5V, as are all other bins. No idea why they set it so high.

The higher memory VRM load triggers some sort of OCP which results in 0V GPU and memory
It's important to note that the controller ISL6534 does NOT actually have OCP or OVP protections. I determined that due to the higher VRM load with faster memory, the 2.8V sometimes spikes to above 3.2V in the nanosecond-range.
Why is that a problem? Because the controller has a internal referance voltage of 3.3V and all VOUT rails MUST be lower than 3.3V, or the controller detects a sort of "short circuit" scenario (might not be strictly accurate but close enough).

By having a good look at the Gigabyte PCX5900 schematics, which use an identical PCB, I was able to reverse-voltmod the memory down to 2.45V which is much closer to the specification and is also plenty to run at 475MHz. The trick here is that the corresponding feedback divider resistor network isn't connected to the FB2 pin, but to the VREF Pin (10) of the controller.

TLDR: Lower memory voltage (complicated), no more protection triggered, card finally able to clock freely.

3D Marks attached. Might be the first PCI-E NV38 card with full 5950 Ultra clockspeeds, except maybe some PCX5950 sample boards.

The attachment IMG_20241110_221217746.jpg is no longer available
The attachment IMG_20241110_221200428.jpg is no longer available
The attachment IMG_20241110_221144613.jpg is no longer available
The attachment PCX5900 475-475.JPG is no longer available
The attachment PCX5900 475-475 01.JPG is no longer available

ToDos left:
Make Voltmod for GPU permanent
Test card in x16 system with faster CPU and proper drivers and compare to 5950U AGP

I'd like to see how far this card can go, any updates? Quadro FX1300 is one of my favorite cards.