Page 5 of 5

Re: How good is the RTX 2080 Ti for Leela?

Posted: Wed Nov 07, 2018 12:37 pm
by Gian-Carlo Pascutto
Another thing to notice in this table is that fp16 numbers tend to be >> 2x fp32 numbers.

This makes me a bit suspicious.

Re: How good is the RTX 2080 Ti for Leela?

Posted: Wed Nov 14, 2018 8:00 pm
by h1a8
I have a question.
I can buy a dual 2080 OC with 8gb GDDR6 each
for a similar price as
A single 2080 ti OC with 11GB GDDR6

Which would give the stronger Leela performance and by how much?

Re: How good is the RTX 2080 Ti for Leela?

Posted: Wed Nov 14, 2018 8:18 pm
by Milos
h1a8 wrote:
Wed Nov 14, 2018 8:00 pm
I have a question.
I can buy a dual 2080 OC with 8gb GDDR6 each
for a similar price as
A single 2080 ti OC with 11GB GDDR6

Which would give the stronger Leela performance and by how much?
Dual 2080 hands down.

Re: How good is the RTX 2080 Ti for Leela?

Posted: Wed Nov 14, 2018 8:20 pm
by Dann Corbit
Milos wrote:
Wed Nov 14, 2018 8:18 pm
h1a8 wrote:
Wed Nov 14, 2018 8:00 pm
I have a question.
I can buy a dual 2080 OC with 8gb GDDR6 each
for a similar price as
A single 2080 ti OC with 11GB GDDR6

Which would give the stronger Leela performance and by how much?
Dual 2080 hands down.
Better have a big power supply.

Re: How good is the RTX 2080 Ti for Leela?

Posted: Thu Nov 15, 2018 6:42 am
by ankan
Gian-Carlo Pascutto wrote:
Wed Nov 07, 2018 12:37 pm
Another thing to notice in this table is that fp16 numbers tend to be >> 2x fp32 numbers.

This makes me a bit suspicious.
That is because we use Tensor Cores for FP16 path. The speedup is only about 3X (and not 8X as raw TFlops nos would suggest) because fp32 path uses Winograd algorithm for convolutions but FP16 path doesn't (not well supported by cudnn).
Right now we don't have a path using FP16 without using tensor math. It should be pretty easy to support - just need to change tensor layout, tensor math option in cudnn and cudnn algorithm selection setting but it would be slower than the current version (with tensor cores enabled).
AFAIK, there is only one Nvidia GPU (P100) supporting high throughput fp16 that doesn't have tensor cores so adding another path just to support it is probably not worth given that P100 is anyway a server only product (Tesla/Quadro).

Re: How good is the RTX 2080 Ti for Leela?

Posted: Tue Nov 20, 2018 9:11 am
by h1a8
Milos wrote:
Wed Nov 14, 2018 8:18 pm
h1a8 wrote:
Wed Nov 14, 2018 8:00 pm
I have a question.
I can buy a dual 2080 OC with 8gb GDDR6 each
for a similar price as
A single 2080 ti OC with 11GB GDDR6

Which would give the stronger Leela performance and by how much?
Dual 2080 hands down.
Thank you! From the spreadsheet,
the rtx 2080 gives about 26000 nps and the rtx 2080 ti gives about 37500 nps. A dual rtx 2080 would give about how many nps?
jpqy wrote:
Thu Oct 25, 2018 9:32 am
In Ipman's Lc0 Benchmark is also RTX 2070 included..2080 & 2080 Ti used same Lc0 v0.18.1
There are the differences bigger..

https://docs.google.com/spreadsheets/d/ ... 1508569046

JP.
I have a few questions about the spreadsheet.

1. Why is the rtx 2080 ti is listed for 3 threads and for 2 threads? Are there two different versions we can buy?

2. Why does the rtx 2070 have 4 threads and not 2? Is that really a dual 2070 setup?

3. Why is the 2 thread rtx 2070 (slightly OC) significantly faster than the 4 thread rtx 2070?

4. Would a dual rtx 2080 setup result in about 52000nps for Leela (since a single 2080 is about 26000nps)?

Re: How good is the RTX 2080 Ti for Leela?

Posted: Tue Nov 20, 2018 3:03 pm
by Robert Pope
The threads refers to how many instances of Lc0 you have running at the same time, nothing to do with different GPU cards. Fast cards don't run at full capacity with just one CPU thread, so you get more output by having two threads sending data to the GPU.

Re: How good is the RTX 2080 Ti for Leela?

Posted: Wed Nov 21, 2018 1:19 pm
by h1a8
Robert Pope wrote:
Tue Nov 20, 2018 3:03 pm
The threads refers to how many instances of Lc0 you have running at the same time, nothing to do with different GPU cards. Fast cards don't run at full capacity with just one CPU thread, so you get more output by having two threads sending data to the GPU.
I don't understand.
So I can buy a rtx 2080ti and make it run 3 threads (instead of 2) and achieve more than 42knps?
My other questions were not answered.


3. Why is the 2 thread rtx 2070 (slightly OC) significantly faster than the 4 thread rtx 2070?

4. Would a dual rtx 2080 setup result in about 52000nps for Leela (since a single 2080 is about 26000nps)?

Re: How good is the RTX 2080 Ti for Leela?

Posted: Wed Nov 21, 2018 1:20 pm
by Milos
h1a8 wrote:
Tue Nov 20, 2018 9:11 am
Milos wrote:
Wed Nov 14, 2018 8:18 pm
h1a8 wrote:
Wed Nov 14, 2018 8:00 pm
I have a question.
I can buy a dual 2080 OC with 8gb GDDR6 each
for a similar price as
A single 2080 ti OC with 11GB GDDR6

Which would give the stronger Leela performance and by how much?
Dual 2080 hands down.
Thank you! From the spreadsheet,
the rtx 2080 gives about 26000 nps and the rtx 2080 ti gives about 37500 nps. A dual rtx 2080 would give about how many nps?
Dual 2080 gives almost twice nps compared to single 2080 (if you set 2x more CPU threads to it with all other parameters being the same).
However, no one know what is the real strength improvement since like with A/B search having it on 2 cores proved 2x nps but not 2x strength improvement.