4 ms·

I'm very surprised that Nvidia chips don't have memory size locked in hardware.

Nextgrid 6y ago
They kind of did - there are resistors on the card's PCB to inform the firmware of how much memory is installed and this operation modified them to match the new amount of memory installed. But beyond that, what would be the reason of spending effort locking this down even more?
- baybal2 6y ago
  Prevent market segmentation bypass.
  - Nextgrid 6y ago
    The effort & equipment required to perform such an upgrade is significant and is very unlikely to be used at scale for defeating market segmentation.
    - baybal2 6y ago
      Well, there is a precedent of graphic cards with modded bios being sold.
      - Red_Leaves_Flyy 6y ago
        Shhh.
    - selfhoster11 6y ago
      I can see that happen as an upgrade service on ebay. It just needs to be priced competitively.
      - gruez 6y ago
        I don't see the economics working out. You're going to be competing against the 3090, which has 24GB of memory and costs 3x as much as a 3070 (MSRP), but also 78% more shaders. Just to match the 3090's compute capacity is going to cost you $1000 (or $891 if you amortize over multiple cards). That leaves you $500 for the memory chips, labor, and equipment. According to [1], the memory chips cost $12/GB. You won't be able to reuse the existing memory chips, so that means you'll need to buy 16GB of chips for each of the cards, which works out to $384. At this point there's only $116 left, and I doubt anyone would go for this considering that the 3090 only costs $116 more, but has more memory (24 GB), and has full warranty. [1] https://www.hardwaretimes.com/doubling-graphics-memory-will-cost-nvidia-and-amd-less-than-12-gb-its-very-likely/ https://www.hardwaretimes.com/doubling-graphics-memory-will-...
        baybal2 6y ago
        3090 is largely sold out despite the ludicrous pricetag
      - bserge 6y ago
        The cost is still going to be high and you risk killing your card. Even if someone offered this kind of service, it would only be taken by a very small number of enthusiasts. So nVidia doesn't care.
        selfhoster11 6y ago
        If someone is skilled enough to work on surface-mount/BGA components, they likely already know what they are doing. Board level repair vs board level upgrade is a very similar process (to the untrained eye such as mine, at least).
- _verandaguy 6y ago
  Economic reasons aside — there's historical precedent for Nvidia creating arbitrary limitations in software (and sometimes in dedicated hardware) for their cards. Famously, many discrete GeForce cards don't play well with virtualized OSes, and there have been some hacky ways to bypass that. This is done because only Quadro cards are blessed by Nvidia to run in "enterprise" environments. Besides this soft lock, there's usually no architectural limitation preventing GeForce cards from working with virtualized OSes if the hypervisor can provide passthrough. All that to say, it wouldn't be the most shocking news in the world if it turned out they prevent this kind of VRAM modding.
  - babypuncher 6y ago
    I'm having trouble imagining a world where the cost of this operation is less than the cost difference between a 3070 and an equivalent 16GB card. Buying the ridiculously overpriced 3090 is still going to be cheaper and not come with any of the associated risk (invalid warranty, potential future driver/bios compatibility problems).
    - _verandaguy 6y ago
      In this specific case, I think it's more about if you can do it rather than if it makes sense. It's a fun challenge, and if you can spare the cash, it does inform the community a bit more about how modern Nvidia cards react to this kind of mod. A big fraction of Hackaday's content is this kind of fun modding that might not make immediate economic sense.
PragmaticPulp 6y ago
It’s likely that nVidia planned for more RAM on these cards early in the development process. Early rumors pointed to 16GB models of the 3070 and 20GB models of the 3080. Due to chip shortages, most likely, we got the smaller amount of RAM on production models. This leads to weird situations where the 3060 has more RAM (12GB) than the 3070 (8GB) and 3080 (10GB). nVidia must use RAM in multiples of the bus width, so the only other option was 16GB for the 3070, which is what this person enabled by swapping chips. nVidia doesn’t care about modders swapping memory chips because it’s not a threat to their sales. I say this as someone who has all of the hot air rework equipment and experience to do this swap: It’s not going to become a common mod. It’s likely not even feasible for someone to offer as a service due to the risk involved in shipping expensive GPUs two ways and then reworking the PCB with hard to source chips that can’t necessarily be guaranteed to work at the exact speeds expected by the drivers. It’s also possible that this mod will be rendered useless if a future driver update introduces different timing parameters, such as when nVidia releases official 16GB 3070 models.
- CivBase 6y ago
  > It’s not going to become a common mod. It’s likely not even feasible for someone to offer as a service due to the risk involved in shipping expensive GPUs two ways and then reworking the PCB with hard to source chips that can’t necessarily be guaranteed to work at the exact speeds expected by the drivers. That's too bad. I prefer to buy a top-end graphics card once every 3 or 4 generations. I was up for a refresh with the 20-series, but the value was pretty bad so I waited. Now the 30-series looks great except for one glaring issue (aside from availability): VRAM capacity. With the imminent release of the direct storage API for Windows, I expect the 30-series cards will show their age much faster than previous generations thanks to their relatively small VRAM capacity compared to Radeon and game consoles.
szczys 6y ago
Doesn't that add manufacturing complexity? If you have multiple firmwares it's possible to flash the wrong one at the factory leading to a lot of pain if they get out to the market before this is discovered (your news 16 GB card is only using 8 GB of memory). This current scheme means there is one firmware to rule them all and the configuration of the hardware itself acts as the selector.
bserge 6y ago
Probably because people don't do this at scale. Fermi and Kepler also used laser cuts near the die to disable cores, some people enabled them with a bit of soldering (not always successful because the cores are often actually non-functional). Not sure about newer ones. Maxwell cards could also be upgraded from 4-6GB to 8GB. This has actually been a cool mod for a long time, earliest I can remember are the ATi Radeon 9x00 and nVidia GeForce 4 Ti cards, people used to double the RAM for fun and profit.

They kind of did - there are resistors on the card's PCB to inform the firmware of how much memory is installed and this operation modified them to match the new amount of memory installed. But beyond that, what would be the reason of spending effort locking this down even more?

baybal2 6y ago
Prevent market segmentation bypass.
- Nextgrid 6y ago
  The effort & equipment required to perform such an upgrade is significant and is very unlikely to be used at scale for defeating market segmentation.
  - baybal2 6y ago
    Well, there is a precedent of graphic cards with modded bios being sold.
    - Red_Leaves_Flyy 6y ago
      Shhh.
  - selfhoster11 6y ago
    I can see that happen as an upgrade service on ebay. It just needs to be priced competitively.
    - gruez 6y ago
      I don't see the economics working out. You're going to be competing against the 3090, which has 24GB of memory and costs 3x as much as a 3070 (MSRP), but also 78% more shaders. Just to match the 3090's compute capacity is going to cost you $1000 (or $891 if you amortize over multiple cards). That leaves you $500 for the memory chips, labor, and equipment. According to [1], the memory chips cost $12/GB. You won't be able to reuse the existing memory chips, so that means you'll need to buy 16GB of chips for each of the cards, which works out to $384. At this point there's only $116 left, and I doubt anyone would go for this considering that the 3090 only costs $116 more, but has more memory (24 GB), and has full warranty. [1] https://www.hardwaretimes.com/doubling-graphics-memory-will-cost-nvidia-and-amd-less-than-12-gb-its-very-likely/ https://www.hardwaretimes.com/doubling-graphics-memory-will-...
      - baybal2 6y ago
        3090 is largely sold out despite the ludicrous pricetag
    - bserge 6y ago
      The cost is still going to be high and you risk killing your card. Even if someone offered this kind of service, it would only be taken by a very small number of enthusiasts. So nVidia doesn't care.
      - selfhoster11 6y ago
        If someone is skilled enough to work on surface-mount/BGA components, they likely already know what they are doing. Board level repair vs board level upgrade is a very similar process (to the untrained eye such as mine, at least).
_verandaguy 6y ago
Economic reasons aside — there's historical precedent for Nvidia creating arbitrary limitations in software (and sometimes in dedicated hardware) for their cards. Famously, many discrete GeForce cards don't play well with virtualized OSes, and there have been some hacky ways to bypass that. This is done because only Quadro cards are blessed by Nvidia to run in "enterprise" environments. Besides this soft lock, there's usually no architectural limitation preventing GeForce cards from working with virtualized OSes if the hypervisor can provide passthrough. All that to say, it wouldn't be the most shocking news in the world if it turned out they prevent this kind of VRAM modding.
- babypuncher 6y ago
  I'm having trouble imagining a world where the cost of this operation is less than the cost difference between a 3070 and an equivalent 16GB card. Buying the ridiculously overpriced 3090 is still going to be cheaper and not come with any of the associated risk (invalid warranty, potential future driver/bios compatibility problems).
  - _verandaguy 6y ago
    In this specific case, I think it's more about if you can do it rather than if it makes sense. It's a fun challenge, and if you can spare the cash, it does inform the community a bit more about how modern Nvidia cards react to this kind of mod. A big fraction of Hackaday's content is this kind of fun modding that might not make immediate economic sense.

PragmaticPulp 6y ago

It’s likely that nVidia planned for more RAM on these cards early in the development process. Early rumors pointed to 16GB models of the 3070 and 20GB models of the 3080. Due to chip shortages, most likely, we got the smaller amount of RAM on production models. This leads to weird situations where the 3060 has more RAM (12GB) than the 3070 (8GB) and 3080 (10GB). nVidia must use RAM in multiples of the bus width, so the only other option was 16GB for the 3070, which is what this person enabled by swapping chips. nVidia doesn’t care about modders swapping memory chips because it’s not a threat to their sales. I say this as someone who has all of the hot air rework equipment and experience to do this swap: It’s not going to become a common mod. It’s likely not even feasible for someone to offer as a service due to the risk involved in shipping expensive GPUs two ways and then reworking the PCB with hard to source chips that can’t necessarily be guaranteed to work at the exact speeds expected by the drivers. It’s also possible that this mod will be rendered useless if a future driver update introduces different timing parameters, such as when nVidia releases official 16GB 3070 models.

CivBase 6y ago
> It’s not going to become a common mod. It’s likely not even feasible for someone to offer as a service due to the risk involved in shipping expensive GPUs two ways and then reworking the PCB with hard to source chips that can’t necessarily be guaranteed to work at the exact speeds expected by the drivers. That's too bad. I prefer to buy a top-end graphics card once every 3 or 4 generations. I was up for a refresh with the 20-series, but the value was pretty bad so I waited. Now the 30-series looks great except for one glaring issue (aside from availability): VRAM capacity. With the imminent release of the direct storage API for Windows, I expect the 30-series cards will show their age much faster than previous generations thanks to their relatively small VRAM capacity compared to Radeon and game consoles.

szczys 6y ago

Doesn't that add manufacturing complexity? If you have multiple firmwares it's possible to flash the wrong one at the factory leading to a lot of pain if they get out to the market before this is discovered (your news 16 GB card is only using 8 GB of memory). This current scheme means there is one firmware to rule them all and the configuration of the hardware itself acts as the selector.

bserge 6y ago

Probably because people don't do this at scale. Fermi and Kepler also used laser cuts near the die to disable cores, some people enabled them with a bit of soldering (not always successful because the cores are often actually non-functional). Not sure about newer ones. Maxwell cards could also be upgraded from 4-6GB to 8GB. This has actually been a cool mod for a long time, earliest I can remember are the ATi Radeon 9x00 and nVidia GeForce 4 Ti cards, people used to double the RAM for fun and profit.