4 ms·

Gamers like to ignore inflation and increasing fab costs and pretend cards should cost the same forever with double performance gains every 1.5 years.

Mountain_Skies 3y ago
For much of tech hardware world, declining costs and increasing performance have been the general trends for as long as most of us have been alive.
- sgarman 3y ago
  Off the top of my head the only thing I can think of that did that was TVs.
  - buzzerbetrayed 3y ago
    Personal computers have plummeted in price over the decades
  - HDThoreaun 3y ago
    CPU's are cheaper adjusted for inflation before even considering performance. Ram and SSDs are much much cheaper than they were 5 years ago.
    - paulmd 3y ago
      RAM was much cheaper 7 years ago than 5 years ago. There was another DRAM cartel action and prices tripled in less than 6 months back in 2017.
  - distances 3y ago
    As far as I can tell computers are still cheaper than they were in the 90s. And that's in absolute numbers, not even accounting for inflation.
solardev 3y ago
Why do GPUs get so much more expensive compared to the rest of the system? CPUs, hard drives, prebuilt systems, laptops etc. haven't gotten ridiculously more expensive. It used to be that $200 would get you a low-range GPU, $400 a midrange, and $600 a pretty darned good one. The 1070 launched at $379 (or $480 in 2023 dollars), the 2070 at $499, the 3070 at $499, then all of a sudden the 4070 now is $599. That's a big difference, even accounting for inflation. The 1080 Ti was $699, an already-unfathomable price back then that cost more than a whole console setup. Today? The 4080 is $1200. That's often more than the rest of the system put together.
- paulmd 3y ago
  It’s because CPUs tend to be fundamentally limited in the ways they can efficiently utilize transistors to scale performance, by things like cache latency, reorder depth, branch prediction, etc. While gpus have always been god’s strongest soldier for putting transistors on silicon scalably. They were the perfect machine for a world where transistors per dollar doubled every 18 months. On the other hand now that it’s stopped, so has raw performance scaling. There was a few generations of cleanup, but you can’t squeeze blood from a stone forever. Maxwell arguably cut too far and pascal had to start adding functionality back. Gpus are like e-cores, large powerful cores mean you get fewer of them and that often works out to a lower PPA. There aren’t many opportunities for cool tricks and the model doesn’t favor using lots of area. The coding is written for extreme parallelism already, so, that’s not a problem, it’s inherent to the platform. Transistor per $ growth hasn’t completely stopped but it’s certainly nowhere near what it was 10 years ago, even if you factor in things like packaging/stacking the total wafer area still runs up a big bill that offset most of the density gains. And that’s what we’ve been seeing over the last 10 years in gpus too. 4070 is about the same area and cutdown as GTX 1070 but wafers cost like 8x as much. What you have to do in this operating regime is find ways to get more performance per transistor - and that’s exactly what Jensen made a big bet on 5 years ago with dlss. 7% more transistors that with DLSS 2.5 give 30% speedup at native quality and 50-70% speedup at iso-quality with FSR2 Quality mode. That’s what the future looks like - rewriting your applications to take advantage of new accelerators that provide large speedups. And the rewriting is very minimal for any application that uses TAAU already. It sucks but if cost/transistor is not going to come down you have to get more out of the transistors you have. Work smarter not harder.
jules22 3y ago
That's fine, and gamers can no longer be expected to upgrade every generation or two, like they used to, if the improvements are going to be marginal between generations. The enthusiast market will just shrink and of course that no longer matters to Nvidia as they are playing in bigger leagues now. GPUs will be bought like TVs. It took 4060, about 6 years to double the Passmark score over 1060.

Mountain_Skies 3y ago

For much of tech hardware world, declining costs and increasing performance have been the general trends for as long as most of us have been alive.

sgarman 3y ago
Off the top of my head the only thing I can think of that did that was TVs.
- buzzerbetrayed 3y ago
  Personal computers have plummeted in price over the decades
- HDThoreaun 3y ago
  CPU's are cheaper adjusted for inflation before even considering performance. Ram and SSDs are much much cheaper than they were 5 years ago.
  - paulmd 3y ago
    RAM was much cheaper 7 years ago than 5 years ago. There was another DRAM cartel action and prices tripled in less than 6 months back in 2017.
- distances 3y ago
  As far as I can tell computers are still cheaper than they were in the 90s. And that's in absolute numbers, not even accounting for inflation.

solardev 3y ago

Why do GPUs get so much more expensive compared to the rest of the system? CPUs, hard drives, prebuilt systems, laptops etc. haven't gotten ridiculously more expensive. It used to be that $200 would get you a low-range GPU, $400 a midrange, and $600 a pretty darned good one. The 1070 launched at $379 (or $480 in 2023 dollars), the 2070 at $499, the 3070 at $499, then all of a sudden the 4070 now is $599. That's a big difference, even accounting for inflation. The 1080 Ti was $699, an already-unfathomable price back then that cost more than a whole console setup. Today? The 4080 is $1200. That's often more than the rest of the system put together.

paulmd 3y ago
It’s because CPUs tend to be fundamentally limited in the ways they can efficiently utilize transistors to scale performance, by things like cache latency, reorder depth, branch prediction, etc. While gpus have always been god’s strongest soldier for putting transistors on silicon scalably. They were the perfect machine for a world where transistors per dollar doubled every 18 months. On the other hand now that it’s stopped, so has raw performance scaling. There was a few generations of cleanup, but you can’t squeeze blood from a stone forever. Maxwell arguably cut too far and pascal had to start adding functionality back. Gpus are like e-cores, large powerful cores mean you get fewer of them and that often works out to a lower PPA. There aren’t many opportunities for cool tricks and the model doesn’t favor using lots of area. The coding is written for extreme parallelism already, so, that’s not a problem, it’s inherent to the platform. Transistor per $ growth hasn’t completely stopped but it’s certainly nowhere near what it was 10 years ago, even if you factor in things like packaging/stacking the total wafer area still runs up a big bill that offset most of the density gains. And that’s what we’ve been seeing over the last 10 years in gpus too. 4070 is about the same area and cutdown as GTX 1070 but wafers cost like 8x as much. What you have to do in this operating regime is find ways to get more performance per transistor - and that’s exactly what Jensen made a big bet on 5 years ago with dlss. 7% more transistors that with DLSS 2.5 give 30% speedup at native quality and 50-70% speedup at iso-quality with FSR2 Quality mode. That’s what the future looks like - rewriting your applications to take advantage of new accelerators that provide large speedups. And the rewriting is very minimal for any application that uses TAAU already. It sucks but if cost/transistor is not going to come down you have to get more out of the transistors you have. Work smarter not harder.

jules22 3y ago

That's fine, and gamers can no longer be expected to upgrade every generation or two, like they used to, if the improvements are going to be marginal between generations. The enthusiast market will just shrink and of course that no longer matters to Nvidia as they are playing in bigger leagues now. GPUs will be bought like TVs. It took 4060, about 6 years to double the Passmark score over 1060.