3 ms·

Those aren't GeForce cards. They don't want their consumer line (GeForce) competing with their datacenter line (Tesla). They previously reduced the FP64/FP16

Those aren't GeForce cards. They don't want their consumer line (GeForce) competing with their datacenter line (Tesla). They previously reduced the FP64/FP16 performance on GeForce cards (making them much worse for neural networks), but it's easier to limit the losses legally than by reducing hardware. Modern games don't want artificially limited ALUs!

Nition 9y ago
Ah sorry, I see, so there's already that split in consumer vs. datacenter and they're now forcing the line to be drawn.
shaklee3 9y ago
This is not technically correct. I think you're referring to cases of Pascal cards where the GeForce version had poor fp64 performance. This is because it's a different asic Rev. Same reason the more expensive p100 Tesla didn't have int8 support -- certainly not because they did it in software. There's no reason a GeForce card for gaming needs high fp64 support. The reason you're seeing it on the Titan v is because it's literally the same asic, but now you're also paying a premium and it's not made for gamers.
- Scaevolus 9y ago
  I'm referring to the GTX 1080 running FP16 at 1/64th the rate of FP32: https://www.anandtech.com/show/10325/the-nvidia-geforce-gtx-1080-and-1070-founders-edition-review/5 https://www.anandtech.com/show/10325/the-nvidia-geforce-gtx-... "GeForce GTX 1080, on the other hand, is not faster at FP16. In fact it’s downright slow. For their consumer cards, NVIDIA has severely limited FP16 CUDA performance. GTX 1080’s FP16 instruction rate is 1/128th its FP32 instruction rate, or after you factor in vec2 packing, the resulting theoretical performance (in FLOPs) is 1/64th the FP32 rate, or about 138 GFLOPs."
  - shaklee3 9y ago
    Right, and I'm saying it uses the GP104 asic, which has that disabled so that other features can be enabled (higher clocks, for example). I don't think you can find and example where the Tesla shares the same chip as the GeForce and something like fp16 is different. I'm sure they do a lot of shady things, but I believe part of this is just chip design tradeoffs.
    - shaklee3 9y ago
      Just to prove the point, the Tesla p40 has the same chip as the GeForce you site, and it too has slower fp16. This is for a $5k card .

Nition 9y ago

Ah sorry, I see, so there's already that split in consumer vs. datacenter and they're now forcing the line to be drawn.

shaklee3 9y ago

This is not technically correct. I think you're referring to cases of Pascal cards where the GeForce version had poor fp64 performance. This is because it's a different asic Rev. Same reason the more expensive p100 Tesla didn't have int8 support -- certainly not because they did it in software. There's no reason a GeForce card for gaming needs high fp64 support. The reason you're seeing it on the Titan v is because it's literally the same asic, but now you're also paying a premium and it's not made for gamers.

Scaevolus 9y ago
I'm referring to the GTX 1080 running FP16 at 1/64th the rate of FP32: https://www.anandtech.com/show/10325/the-nvidia-geforce-gtx-1080-and-1070-founders-edition-review/5 https://www.anandtech.com/show/10325/the-nvidia-geforce-gtx-... "GeForce GTX 1080, on the other hand, is not faster at FP16. In fact it’s downright slow. For their consumer cards, NVIDIA has severely limited FP16 CUDA performance. GTX 1080’s FP16 instruction rate is 1/128th its FP32 instruction rate, or after you factor in vec2 packing, the resulting theoretical performance (in FLOPs) is 1/64th the FP32 rate, or about 138 GFLOPs."
- shaklee3 9y ago
  Right, and I'm saying it uses the GP104 asic, which has that disabled so that other features can be enabled (higher clocks, for example). I don't think you can find and example where the Tesla shares the same chip as the GeForce and something like fp16 is different. I'm sure they do a lot of shady things, but I believe part of this is just chip design tradeoffs.
  - shaklee3 9y ago
    Just to prove the point, the Tesla p40 has the same chip as the GeForce you site, and it too has slower fp16. This is for a $5k card .