about the term of equal hardware

[j.korhonen]

Tell me more about Zeta GPU?

In short, it is a port of Zeta Dva engine to OpenCL to run on a GPU.

https://www.chessprogramming.org/Zeta

https://zeta-chess.app26.de/

--
Srdja

classic parallel Alpha-Beta on GPU and open source. WOW. Why you not in TCEC?

~2200 elo for now

And what? TCEC Season 15. Division 4a. Igel 1.2 2386 Elo
Such a unique engine as Zeta should be play in the TCEC

Igel is "at least" 2800-2900 on TCEC hardware.

After the completion of Division 4a live rating of Igel 1.2 is 3232 Elo on TCEC hardware https://cd.tcecbeta.club/archive.html?s ... =4a&game=1 (Tab Standings )

[Guenther]

Tell me more about Zeta GPU?

In short, it is a port of Zeta Dva engine to OpenCL to run on a GPU.

https://www.chessprogramming.org/Zeta

https://zeta-chess.app26.de/

--
Srdja

classic parallel Alpha-Beta on GPU and open source. WOW. Why you not in TCEC?

~2200 elo for now

And what? TCEC Season 15. Division 4a. Igel 1.2 2386 Elo
Such a unique engine as Zeta should be play in the TCEC

Igel is "at least" 2800-2900 on TCEC hardware.

After the completion of Division 4a live rating of Igel 1.2 is 3232 Elo on TCEC hardware https://cd.tcecbeta.club/archive.html?s ... =4a&game=1 (Tab Standings )

No you read the table wrong. The estimate of TCEC before the Div 4a was 3232, which was of course wrong by several hundreds of rating points
and also for most other programs of those 10 participants.

Code: Select all

 1      Nemorino 5.08           18      15      0       13 [7/6]        1 [1/0]         121.50  2986    33 [3019]
 2      RubiChess 1.3.1         18      12.5    0       10 [5/5]        3 [1/2]         92.25   3000     7 [3007]
 3      Winter 0.5              18      12.5    0       11 [6/5]        4 [1/3]         87.75   2584    86 [2670]
 4      Rodent III 0.276        18      11      0       8 [5/3]         4 [2/2]         83.50   2713    49 [2762]
 5      The Baron 3.44.1        18      10.5    0       8 [6/2]         5 [2/3]         71.25   2816    23 [2839]
 6      Tucano 7.07             18      9.5     0       7 [3/4]         6 [3/3]         65.25   2685    40 [2725]
 7      Cheese 2.0              18      8       0       6 [4/2]         8 [2/6]         63.25   2700    22 [2722]
 8      Minic 0.42              18      5       0       4 [3/1]         12 [5/7]        30.25   3000   -63 [2937]
 9      Topple 0.4.0            18      3.5     0       3 [3/0]         14 [5/9]        21.00   3000   -78 [2922]
10      Igel 1.2                18      2.5     0       1 [1/0]         14 [6/8]        17.50   3372  -140 [3232]

The list below is an excerpt of identical or very similar versions from CCRL 40/40.
Note that I only selected 1cpu versions for all (CCRL has only 1 and 4 anyway), but that doesn' matter
much for the order of programs, which is decisive. Adding a fantasy number to each program due to
TCEC hardware won't change it...

Even with 18 games only, Igel was clearly last, while seeded 'by lottery' clearly for #1 ;-)

Code: Select all

CCRL 40/40

Nemorino 5.00 64-bit    3087    +16     -16     46.1%   +30.8   49.5%   1195
RubiChess 1.3 64-bit    2973    +26     -26     49.0%   +5.0    41.2%   464
Winter 0.5 64-bit       2900    +33     -33     53.5%   -21.3   47.5%   282
The Baron 3.44 64-bit   2855    +17     -17     49.4%   +3.5    44.6%   1128
Cheese 2.0 64-bit       2746    +18     -18     51.3%   -8.6    40.3%   1001
Igel 1.2 64-bit         2627    +20     -20     48.9%   +6.0    29.4%   880

[voffka]

I've been inspired by this discussion to test Igel 1.4.1 (~2660) on a hardware that is close to TCEC and after executing 226 CCRL 40/40 games on the following machine:

Intel Core i7-6950X CPU @ 3.00GHz (10 cores, 20 threads)
128GB of RAM

In the following configuration:

Igel 1.3 (~2660)
Hash=16 GB
Threads=19
Tablebases: no
Book=Perfect2017.bin, 12 moves

against

GreKo 2018 (~2710)
Hash=1 GB (max for GreKo)
Threads=1 (max for GreKo since smp is not supported)
Tablebases: no
Book=Perfect2017.bin, 12 moves

I got pretty much a draw score.

P.S. I tried to attach a pgn file of the games but it tells me the file type is invalid???

[Graham Banks]

Prior to the start of TCEC, I gave Anton the following rough estimations of engine strength under CCRL testing:

Jumbo 0.6.96 (Sven Schule) - 2500
Bagatur 1.6b (Krasimir Topchiyski) - 2500
Topple 0.3.5 (Vincent Tang) - 2700
Marvin 3.2.0 (Martin Danielsson) - 2850
Igel 1.2 (Volodymyr Shcherbyna) - 2600
Cheese 2.0 (Patrice Duhamel) - 2750
RubiChess 1.3 (Andreas Matthies) - 3000
The Baron 3.44.1 (Richard Pijl) - 2850
Arminius 2018-11-23 (Volker Annuss) - 2700
Monolith 1.02 (Jonathan Mayr) - 2800
Minic 0.40 (Vivian Clauzon) - 2500

[j.korhonen]

Okay, another try,

let us assume the best software is those which makes best use of the transistors
given on a specific chip.

So let us compare the transistor count and the clock frequency of TCEC hardware:

Nvidia RTX 2080 TI:
18.6 billion transistors clocked with max 1.545 GHz
Nvidia RTX 2080:
13.6 billion transistors clocked with max 1.710 GHz

2*Intel Xeon E5-2699v4:
2*7.2 billion transistors clocked with max 3.6 GHz


Even with the best black magic bit twiddling hacks you will be limited by
transistor count and frequency, now do the maths and tell me where the
hardware advantage is...

--
Srdja

Nvidia RTX 2080 TI: 250W
+
Nvidia RTX 2080: 215W
= 465W

Intel Xeon E5-2699v4: 145 W
+
Intel Xeon E5-2699v4: 145 W
= 290W

[Dann Corbit]

P.S. I tried to attach a pgn file of the games but it tells me the file type is invalid???

Try putting them in a zip archive or a 7zip archive.

I guess that they are trying to save space by making us compress stuff.

You can attach archives and images. Not sure what else.

[MikeGL]

Okay, another try,

let us assume the best software is those which makes best use of the transistors
given on a specific chip.

So let us compare the transistor count and the clock frequency of TCEC hardware:

Nvidia RTX 2080 TI:
18.6 billion transistors clocked with max 1.545 GHz
Nvidia RTX 2080:
13.6 billion transistors clocked with max 1.710 GHz

2*Intel Xeon E5-2699v4:
2*7.2 billion transistors clocked with max 3.6 GHz


Even with the best black magic bit twiddling hacks you will be limited by
transistor count and frequency, now do the maths and tell me where the
hardware advantage is...

--
Srdja

Nvidia RTX 2080 TI: 250W
+
Nvidia RTX 2080: 215W
= 465W

Intel Xeon E5-2699v4: 145 W
+
Intel Xeon E5-2699v4: 145 W
= 290W

In that case, (2) Intel Xeon E5-2699v4 (290W) vs (1) Nvidia RTX 2080 TI (250W) would be an interesting match with a bit of 40W disadvantage for the NN engine.

[Dann Corbit]

We can measure in watts.
We can measure in dollars.
We can measure in flops.

But we will always be comparing apples to oranges.
GPUs give enormous floating point operation counts.
GPUs are really, really cheap for the compute power you get.
GPUs use a lot of watts.
But these two systems play chess very, very differently.

I think that whatever it was they did for TCEC was pretty good.
The reason I say that is that the strength of LC0 and Stockfish was very similar.
So the contest was competitive.

[smatovic]

We can measure in watts.
We can measure in dollars.
We can measure in flop
But we will always be comparing apples to oranges.
GPUs give enormous floating point operation counts.
GPUs are really, really cheap for the compute power you get.
GPUs use a lot of watts.
But these two systems play chess very, very differently.

I think that whatever it was they did for TCEC was pretty good.
The reason I say that is that the strength of LC0 and Stockfish was very similar.
So the contest was competitive.

+1

...LC0 uses GPUs as neural network accelerators, simply cos they fit well for
this task and are relatively cheap for their performance. If in future there is
an alternative technique available to the end user, like TPUs, FPGAs, or even
Quantum Machines, then LC0 will use them too, regardless of how they compare
to CPUs.

--
Srdja

[smatovic]

Okay, another try,

let us assume the best software is those which makes best use of the transistors
given on a specific chip.

So let us compare the transistor count and the clock frequency of TCEC hardware:

Nvidia RTX 2080 TI:
18.6 billion transistors clocked with max 1.545 GHz
Nvidia RTX 2080:
13.6 billion transistors clocked with max 1.710 GHz

2*Intel Xeon E5-2699v4:
2*7.2 billion transistors clocked with max 3.6 GHz


Even with the best black magic bit twiddling hacks you will be limited by
transistor count and frequency, now do the maths and tell me where the
hardware advantage is...

--
Srdja

Nvidia RTX 2080 TI: 250W
+
Nvidia RTX 2080: 215W
= 465W

Intel Xeon E5-2699v4: 145 W
+
Intel Xeon E5-2699v4: 145 W
= 290W

Watts are imo not appropriate for comparison, there are differences in
fabrication process for different architectures that causes different
consumptions...i am not that into fabs, but Nvidia RTX is produced in TSMC 12FFN
and the Xeons are produces in Intels 14 nm process...maybe Milos could clarify

--
Srdja