3 ms·

Disclosure: I very briefly worked for a competitor. A) Drivers reveal a surprising amount of "secret sauce" for a GPU. The "public-facing" interface to a GPU

Disclosure: I very briefly worked for a competitor. A) Drivers reveal a surprising amount of "secret sauce" for a GPU. The "public-facing" interface to a GPU is something like HLSL or GLSL, but the real interface is some proprietary ISA. The manufacturers are slowly moving in this area, but the culture of the ISA being proprietary remains. That leads to things like implementation details getting inserted into ISA docs, or code comments, or driver architecture designs, because they're for internal use only anyway. This philosophy is slowly changing, but it's still the case quite a bit. B) There's much less difference than you would expect between a $500 gaming GPU and a $2,000 workstation GPU. It's totally plausible that you are one FOSS driver away from a much more expensive GPU. That is a problem for everybody, including the gamers who want to keep buying $500 GPUs. C) There is a VERY complicated balance of power between OEMs (who control the purse), Intel (who sometimes competes and sometimes complements the video card manufacturers) and Microsoft (for better or for worse, a de facto standards body) and increasingly, new kinds of competitors like PowerVR. So this is just a hypothetical, but one way it could go down is like this: NVidia decides to fully embrace FOSS. Microsoft gets pissy and tries to write them out of the standard. NVidia forced to lower prices to get PC makers to buy. Apple swoops in and places a huge order of the now-cheap parts. Intel gets pissy that Apple is shipping too many Nvidia GPUs per Intel GPU, raises Apple's CPU pricing to cover the loss. The price of Macs goes up driving some customers to iPads. PowerVR's royalty fees double, and since they don't have to manage any inventory, they are increasingly more attractive to investors than traditional manufacturers like NVidia. NVidia's stock price falls, and some C-level forgoes a vacation in Europe.

Dylan16807 13y ago
> B) There's much less difference than you would expect between a $500 gaming GPU and a $2,000 workstation GPU. It's totally plausible that you are one FOSS driver away from a much more expensive GPU. That is a problem for everybody, including the gamers who want to keep buying $500 GPUs. I wasn't aware that there was a difference at all. Haven't there been cases of changing a single resistor to switch the card's mode?
- H3g3m0n 13y ago
  That is just how chip manufacturing works. In theory you make 1 model of chip. You manufacture them. You test them. What clock speeds are they stable at. Are any of the cores faulty. Many (probably most) will have defects. Some of them will be duds and must be thrown The bits with the defects are switched off and resold as lower end hardware. The common less powerful failure cases cost less and the rare more powerful successes cost more. The high end stuff sells for more than the cost of manufacturing and the low end stuff goes for less. They offset each other. So when people force a model of GPU to preform higher they are basically ignoring the Q.C. and are going to be dealing with undefined behaviour. Did the card get reduced just because it produced more heat at the higher levels (in which case a highend heatsint might counteract the issue) or because there is a core that returns faulty data. That's only the theory though. In reality it might be artificially screwed with to increase profits. Maybe there are many more successes than the desired price level matches. Maybe they are selling more cheaper stuff so they turn off the bits of the chips even though they passed the QC. There are only 2 companies selling gaming GPUs so there could be an artificial duopoly. Nvidia only have to be price competitive with AMD.
  - Dylan16807 13y ago
    That's binning, sure. But I thought pro chips actually tended to be clocked a bit lower than gaming chips. There's no 'ignoring QC' there if you trick the board, quite the opposite.
    - exDM69 13y ago
      > But I thought pro chips actually tended to be clocked a bit lower than gaming chips. There's no 'ignoring QC' there if you trick the board, quite the opposite. Yes, the pro gpus have different clocks, ECC memory and there may be other differences in the board as well.
- brymaster 13y ago
  Here's a great example: Resistor hack turns a Nvidia GTX690 into a Quadro K5000 or Tesla K10 https://news.ycombinator.com/item?id=5398555 https://news.ycombinator.com/item?id=5398555

Dylan16807 13y ago

> B) There's much less difference than you would expect between a $500 gaming GPU and a $2,000 workstation GPU. It's totally plausible that you are one FOSS driver away from a much more expensive GPU. That is a problem for everybody, including the gamers who want to keep buying $500 GPUs. I wasn't aware that there was a difference at all. Haven't there been cases of changing a single resistor to switch the card's mode?

H3g3m0n 13y ago
That is just how chip manufacturing works. In theory you make 1 model of chip. You manufacture them. You test them. What clock speeds are they stable at. Are any of the cores faulty. Many (probably most) will have defects. Some of them will be duds and must be thrown The bits with the defects are switched off and resold as lower end hardware. The common less powerful failure cases cost less and the rare more powerful successes cost more. The high end stuff sells for more than the cost of manufacturing and the low end stuff goes for less. They offset each other. So when people force a model of GPU to preform higher they are basically ignoring the Q.C. and are going to be dealing with undefined behaviour. Did the card get reduced just because it produced more heat at the higher levels (in which case a highend heatsint might counteract the issue) or because there is a core that returns faulty data. That's only the theory though. In reality it might be artificially screwed with to increase profits. Maybe there are many more successes than the desired price level matches. Maybe they are selling more cheaper stuff so they turn off the bits of the chips even though they passed the QC. There are only 2 companies selling gaming GPUs so there could be an artificial duopoly. Nvidia only have to be price competitive with AMD.
- Dylan16807 13y ago
  That's binning, sure. But I thought pro chips actually tended to be clocked a bit lower than gaming chips. There's no 'ignoring QC' there if you trick the board, quite the opposite.
  - exDM69 13y ago
    > But I thought pro chips actually tended to be clocked a bit lower than gaming chips. There's no 'ignoring QC' there if you trick the board, quite the opposite. Yes, the pro gpus have different clocks, ECC memory and there may be other differences in the board as well.
brymaster 13y ago
Here's a great example: Resistor hack turns a Nvidia GTX690 into a Quadro K5000 or Tesla K10 https://news.ycombinator.com/item?id=5398555 https://news.ycombinator.com/item?id=5398555