CPU vs GPU...fair play...keep it simple

[Hai]

best cpu = best gpu

Intel Xeon Platinum = Nvidia Tesla V100

AMD Threadripper = Nvidia Titan V

AMD Ryzen 7 = Nvidia 1080 Ti

Now do some matches and tournaments.

[Milos]

best cpu = best gpu

Intel Xeon Platinum = Nvidia Tesla V100

AMD Threadripper = Nvidia Titan V

AMD Ryzen 7 = Nvidia 1080 Ti

Now do some matches and tournaments.

You got it all mixed up. Tesla V100 is identical in performance to Titan V when running Lc0 despite 3x difference in price. Even worse is Xeon Platinum 8180, no extra performance vs 2990 and 6x the price.
Threadripper is 3 times cheaper than Titan V
Ryzen 7 is 3 times cheaper than 1080Ti.
If you really wanna realistic price comparison it is:
AMD Threadripper 2990x vs 2x1080Ti
AMD Threadripper 1950x vs 1080Ti
AMD Threadripper 1920x vs 1080
AMD Ryzen 7 1800x vs 1070
AMD Ryzen 7 1700 vs 1060

[Dann Corbit]

best cpu = best gpu

Intel Xeon Platinum = Nvidia Tesla V100

AMD Threadripper = Nvidia Titan V

AMD Ryzen 7 = Nvidia 1080 Ti

Now do some matches and tournaments.

You got it all mixed up. Tesla V100 is identical in performance to Titan V when running Lc0 despite 3x difference in price. Even worse is Xeon Platinum 8180, no extra performance vs 2990 and 6x the price.
Threadripper is 3 times cheaper than Titan V
Ryzen 7 is 3 times cheaper than 1080Ti.
If you really wanna realistic price comparison it is:
AMD Threadripper 2990x vs 2x1080Ti
AMD Threadripper 1950x vs 1080Ti
AMD Threadripper 1920x vs 1080
AMD Ryzen 7 1800x vs 1070
AMD Ryzen 7 1700 vs 1060

That's a pretty good assessment. I agree with you.

From IPMAN's chess site,
http://www.ipmanchess.yolasite.com/amd- ... -bench.php

This setup:

Code: Select all

Nodes/second  CPU                                      Cores/Threads            BMI2/popcnt    Member 
431.403.814    8x Intel Xeon Platinum 8168    8x24c/48t384threads  BMI2               noobpwnftw

Just the CPUs are (approximately) $6200 each if you do not buy them used. Hence:
50K just for the CPUs. We have not even added disk or RAM, etc.
I know that he as more than a terabyte of RAM also, because it is necessary to do the 7 man files.
I guess that his system was well over $100K in total.

Hard to say what would be the equivalent.

Maybe one of these:
https://www.thinkmate.com/system/supers ... 28g2-fc0pt+
with 12 x 1080Ti and the other high end stuff you need in a machine like that to coordinate the calculation.
They want to sell you those Pascal cards, but I guess (at least for chess) that would be a stupid buy.

[Milos]

That's a pretty good assessment. I agree with you.

From IPMAN's chess site,
http://www.ipmanchess.yolasite.com/amd- ... -bench.php

This setup:

Code: Select all

Nodes/second  CPU                                      Cores/Threads            BMI2/popcnt    Member 
431.403.814    8x Intel Xeon Platinum 8168    8x24c/48t384threads  BMI2               noobpwnftw

Just the CPUs are (approximately) $6200 each if you do not buy them used. Hence:
50K just for the CPUs. We have not even added disk or RAM, etc.
I know that he as more than a terabyte of RAM also, because it is necessary to do the 7 man files.
I guess that his system was well over $100K in total.

Hard to say what would be the equivalent.

Maybe one of these:
https://www.thinkmate.com/system/supers ... 28g2-fc0pt+
with 12 x 1080Ti and the other high end stuff you need in a machine like that to coordinate the calculation.
They want to sell you those Pascal cards, but I guess (at least for chess) that would be a stupid buy.

That's easy, DGX station with 4 V100 would be kind of equivalent . It is 70k$.
A step further would be the new (Volta based) DGX 1 with 8 V100. That one is 150k$.
If you really want the ultimate NN machine that would certainly be DGX 2 with 16 V100. But that one is 400k$ .

Btw. noobpwnftw's machine is nowhere near 100k. You get discount when you buy eight of Intel Xeon Platinum 8168 and Intel recommended price was 5890$ which means almost certainly he got them cheaper than that. I would say around 45k for CPUs. 1TB RAM is not that much either, one year ago when he was buying it was around 7$ per GB, so only 7k$ for 1TB. So if you include all stuff his machine was maximum 60k$.

[jdart]

I don't know how NNs scale but for a/b searchers the branching factor is typically 3 or more, so doubling speed doesn't quite get you an extra ply of depth, and that is not a lot of strength increase if you're already hitting pretty high depths.

My take is TCEC is interesting for the higher end than normal hardware, but if you want to know what a realistic performance might be on hardware most people will actually run, match a $1000 CPU (say i9-7940X) with a $1000 graphics card.

However, IMO there is no magic "fair" point to run a competition between a GPU based engine and a CPU based engine. They are different beasts. If you want LC0 to beat Stockfish, you can throw hardware at it until it does that; but that doesn't mean it's a level playing field. What is level is always going to be subjective.

--Jon

[Milos]

I don't know how NNs scale but for a/b searchers the branching factor is typically 3 or more, so doubling speed doesn't quite get you an extra ply of depth, and that is not a lot of strength increase if you're already hitting pretty high depths.

Maybe for Arasan EBF is 3, for SF is 1.8 only .
Don't worry much about MCTS with averaging (A0/Lc0 type) it starts to scale much worse (I mean much worse than A/B) after 1M nodes per move.

However, IMO there is no magic "fair" point to run a competition between a GPU based engine and a CPU based engine. They are different beasts. If you want LC0 to beat Stockfish, you can throw hardware at it until it does that; but that doesn't mean it's a level playing field. What is level is always going to be subjective.

Well if you really wanna beat SF (Droidfish) playing on a cellphone (like newest Iphone or Samsung S9) not even 1080Ti is enough, one needs something stronger like V100 .

[frankp]

I don't know how NNs scale but for a/b searchers the branching factor is typically 3 or more, so doubling speed doesn't quite get you an extra ply of depth, and that is not a lot of strength increase if you're already hitting pretty high depths.

My take is TCEC is interesting for the higher end than normal hardware, but if you want to know what a realistic performance might be on hardware most people will actually run, match a $1000 CPU (say i9-7940X) with a $1000 graphics card.

However, IMO there is no magic "fair" point to run a competition between a GPU based engine and a CPU based engine. They are different beasts. If you want LC0 to beat Stockfish, you can throw hardware at it until it does that; but that doesn't mean it's a level playing field. What is level is always going to be subjective.

--Jon

I agree with your main point. Define 'fair' then you can answer the question. And probably first define 'fair' according to your prejudice and preferred outcome. This discussion seems a version of an eternal debate: 8bit/16/32/64 bit, 1CPU/multi-core CPU, mainframe/desktop etc. There always seems to be two ends of the spectrum: what is the best chess playing entity; and the best that I can run on my 'home pc'.

It seems reasonable to suppose that the 'chips' to run AI code fast will become more common powerful and cheaper in the future. The first steps seem to be being taken now. Similar to mainframes when Bob started and desktop cpu of today.

Congratulations on your promotion btw. Sure was an eventual competition to watch. Arasan losing near the end, certainly added drama. I think the fanbois (on both sides) had already massed with pitchforks in hand.

[modolief]

Equal power consumption. That is by far my number one criteria. After that would be memory; each machine should get the same amount of memory.

I think eventually hardware prices, which are driven by marketing, tend to level out. So I think CPU vs GPU should be judged by watts drained out of the wall per unit of time, all other things being equal.

[megamau]

If you really wanna realistic price comparison it is:
AMD Threadripper 2990x vs 2x1080Ti
AMD Threadripper 1950x vs 1080Ti
AMD Threadripper 1920x vs 1080
AMD Ryzen 7 1800x vs 1070
AMD Ryzen 7 1700 vs 1060

This is exactly what I was trying to find out.
Could you add the prices and estimated nodes for Stockfish and Leela ?

[Werewolf]

Ryzen 7 is 3 times cheaper than 1080Ti.
If you really wanna realistic price comparison it is:
AMD Threadripper 2990x vs 2x1080Ti
AMD Threadripper 1950x vs 1080Ti
AMD Threadripper 1920x vs 1080
AMD Ryzen 7 1800x vs 1070
AMD Ryzen 7 1700 vs 1060

So what CPU would be roughly equal to a Titan V ?