Soon nvidia line will get updated for RTX 20xx (which should cost about the same as 10xx cards today), so I am wondering how much better for example 2080Ti will be compared to 1080Ti. I.e will leela be able to take advantage of the tensor cores or (maybe) fp16 computation for playing strength? Or will the improvement for leela consist mainly from higher clocks / more CUDA cores? Does anyone have any idea?
CPU vs GPU...fair play...keep it simple
Moderators: hgm, Rebel, chrisw
-
- Posts: 52
- Joined: Sat Mar 24, 2018 4:18 pm
Re: CPU vs GPU...fair play...keep it simple
-
- Posts: 1796
- Joined: Thu Sep 18, 2008 10:24 pm
Re: CPU vs GPU...fair play...keep it simple
Well apparently Lc0 can already take advantage of half precision. But I'm not sure if the new gaming cards will support it very well. The fp16 of the current gen is a lot slower than single precision. But cards like Titan show a 2x speedup.mirek wrote: ↑Sat Aug 18, 2018 7:59 pmSoon nvidia line will get updated for RTX 20xx (which should cost about the same as 10xx cards today), so I am wondering how much better for example 2080Ti will be compared to 1080Ti. I.e will leela be able to take advantage of the tensor cores or (maybe) fp16 computation for playing strength? Or will the improvement for leela consist mainly from higher clocks / more CUDA cores? Does anyone have any idea?
-
- Posts: 1796
- Joined: Thu Sep 18, 2008 10:24 pm
Re: CPU vs GPU...fair play...keep it simple
Well apparently Lc0 can already take advantage of half precision. But I'm not sure if the new gaming cards will support it very well. The fp16 of the current gen is a lot slower than single precision. But cards like Titan V show a 2x speedup.mirek wrote: ↑Sat Aug 18, 2018 7:59 pmSoon nvidia line will get updated for RTX 20xx (which should cost about the same as 10xx cards today), so I am wondering how much better for example 2080Ti will be compared to 1080Ti. I.e will leela be able to take advantage of the tensor cores or (maybe) fp16 computation for playing strength? Or will the improvement for leela consist mainly from higher clocks / more CUDA cores? Does anyone have any idea?
-
- Posts: 4190
- Joined: Wed Nov 25, 2009 1:47 am
Re: CPU vs GPU...fair play...keep it simple
Consumer(gaming) cards will never have real fp16 support or tensor cores. NVIDIA is not crazy to ruin themselves the chance to earn more by selling "AI tergeted" cards for much more money.mirek wrote: ↑Sat Aug 18, 2018 7:59 pmSoon nvidia line will get updated for RTX 20xx (which should cost about the same as 10xx cards today), so I am wondering how much better for example 2080Ti will be compared to 1080Ti. I.e will leela be able to take advantage of the tensor cores or (maybe) fp16 computation for playing strength? Or will the improvement for leela consist mainly from higher clocks / more CUDA cores? Does anyone have any idea?
For consumer cards 20xx will be roughly 15% faster than 10xx. That is the same kind of speedup compared to difference between Ryzen/Threadripper 1xxx and 2xxx. So difference between price optimal CPU and GPU machines will remain the same.
-
- Posts: 52
- Joined: Sat Mar 24, 2018 4:18 pm
Re: CPU vs GPU...fair play...keep it simple
Makes sense that they are not crazy to compete with their enterprise grade cards. At the same time though they have been pushing some AI based image post-processing - of which the most important is the AI based denoising of ray traced scenes. And since ray tracing should be a feature of the new consumer cards (otherwise why to call them RTX - and the RTX naming instead of GTX seems very strongly hinted atm) therefore I was thinking that from this point of view it makes sense for them to have at least some of the tensor cores unlocked even on the consumer grade hardware.Milos wrote: ↑Sat Aug 18, 2018 10:21 pmConsumer(gaming) cards will never have real fp16 support or tensor cores. NVIDIA is not crazy to ruin themselves the chance to earn more by selling "AI tergeted" cards for much more money.mirek wrote: ↑Sat Aug 18, 2018 7:59 pmSoon nvidia line will get updated for RTX 20xx (which should cost about the same as 10xx cards today), so I am wondering how much better for example 2080Ti will be compared to 1080Ti. I.e will leela be able to take advantage of the tensor cores or (maybe) fp16 computation for playing strength? Or will the improvement for leela consist mainly from higher clocks / more CUDA cores? Does anyone have any idea?
For consumer cards 20xx will be roughly 15% faster than 10xx. That is the same kind of speedup compared to difference between Ryzen/Threadripper 1xxx and 2xxx. So difference between price optimal CPU and GPU machines will remain the same.
I guess we will know for sure after 20xx series reveal on Monday, for now I take it that the plain 15% speedup to 1080Ti (for lc0) may be the most likely thing to happen but I think that the possibility of the speed up being more serious thanks to some of the tensor cores being unlocked isn't completely off the table yet.