Is disabling a select VRAM channel still a thing?

[Macro Device]

Some time ago, one could remove a VRAM module from their GPU and still have a working GPU, albeit with hindered VRAM bandwidth and amount. Can you still do that to recently made GPUs?

Let's say we don't care if it's moronic. I'm only interested in ifs and hows.

 

[lexluthermiester]

Interesting idea. No clue if it'll work.

 

[eidairaman1]

The bios has been fudged in the past to claim a card has 4GB but physically only has 1 or 2 GB, the damn fake 1060s and 960s, 760s etc were deceptive that way, so in that situatuon games would have an error or crash.

 

[FreedomEclipse]

If i was taking a guess. I has to be coded into the bios.

I remember in the early 90s there were some Matrox Millenniums (G200) & ATi Rage cards that you could buy extra ram for, install yourself and it would work perfectly fine as they were designed with that feature and function in mind. That died with those generations/eras of cards.

 

[Macro Device]

The bios has been fudged in the past to claim a card has 4GB but physically only has 1 or 2 GB, the damn fake 1060s and 960s, 760s etc were deceptive that way, so in that situatuon games would have an error or crash.

No, I'm not talking fake GPUs, I'm talking real GPUs like some real GTX 780 Ti with 3 GB VRAM for example that has one VRAM chip broken/inaccessible and after some fiddling, the owner just disables it and runs a 2.5 GB 780 Ti. Lower FPS in games, sure, but at least it now works.

Not sure if that was possible on this generation, yet I do remember it being possible during this era or perhaps a couple generations earlier.

My interest is 5070 Ti. It has 70 SMs which is roughly 15% better than 4070 Ti non-Super. Maybe 20. No idea. I wanna see how meaningful this additional bandwidth really is for Blackwell and for that to be more informative, I would've loved to have an option to temporarily disable 4 GB of VRAM.

 

[eidairaman1]

If i was taking a guess. I has to be coded into the bios.

I remember in the early 90s there were some Matrox Millenniums (G200) & ATi Rage cards that you could buy extra ram for, install yourself and it would work perfectly fine as they were designed with that feature and function in mind. That died with those generations/eras of cards.

It still applies today for bios coding.

Those cards you mentioned had either a isa like blade or 26 pin header for vesa feature/module connector.

 

[AusWolf]

I didn't know this was ever a thing.

Isn't the VRAM tied to the TMUs somehow? If I remember right, that's why Nvidia got themselves into trouble with the 970 - they disabled some TMUs on the chip, but that also disabled a portion of a memory channel that had to share bandwidth with another channel or something, effectively dropping bandwidth significantly on 0.5 out of the 4 GB VRAM.

 

[StViolenceDay]

The locked VBIOS editing limits this.
So it was possible only up to GTX 7xx generation - https://www.techpowerup.com/forums/...ifacts-code-43-by-disabling-bad-parts.306043/

But for NVIDIA itself it's still a thing.
GPU chips of RTX3060 12GB and RTX3060 8GB are 100% identical.

They were manufactured before the NVIDIA's dicision to "make 3060 less attractive than 4060". So they tell partners to produce newer boards with a moved resistor on the board to flip 0x0040 configurable bit in the DeviceID; Then issued channel-disabling VBIOS for RTX 3060, leaving 8GB out of 12GB with this new DeviceID.

For AMD its even simpler, but also limited by VBIOS locking.
Flashing RX6700 non-XT VBIOS onto RX6700XT board - reduces memory from 12GB to 10GB