Not logged inRybka Chess Community Forum
Up Topic The Rybka Lounge / Computer Chess / Fake computers in Chessbase Engine Cloud
- - By Lukas Cimiotti (Bronze) Date 2021-03-27 12:35 Edited 2021-03-27 15:38 Upvotes 1
In the Chessbase Engine Cloud a provider who calls himself "amd 128 core" is offering big computers:

Hardware and MN/s look impressive, but you won't get what you pay for. I made some video clips of his engines - please watch time to depth

This is his Stockfish 13  256 core cluster - after a minute it reaches depth 39:
http://www.abdul-h.de/Clips/Fake_Cluster.mp4
This is his Stockfish 13  128 core Epyc server - after a minute it reaches depth 37:
http://www.abdul-h.de/Clips/Fake_128_core.mp4
This is one of my cpomuters - a 64 core Threadripper 3990X running at 3.2 GHz  - after a minute it reaches depth 45:
http://www.abdul-h.de/Clips/Real_TR3990X.mp4

This is his lc0  2x RTX 3090 - after a minute it reaches depth 25:
http://www.abdul-h.de/Clips/Fake_2x_RTX3090.mp4
This is one of my cpomuters: 2x RTX 3080 - after a minute it reaches depth 30:
http://www.abdul-h.de/Clips/Real_2xRTX3080.mp4
Parent - By RFK (Gold) Date 2021-03-27 22:06 Edited 2021-03-27 22:12 Upvotes 1
That's interesting! I'm right behind you with an i9-9940x- depth 46 within one minute. That should the guy with the 128-core dude feel a little funny!
Addendum: I'm not overclocking- it probably is running on turbo? I forgot to look. I got out of overclocking these days!
Parent - - By Labyrinth (*****) Date 2021-03-28 03:05
Could it be a scaling issue or latency due to multiple sockets? This would also apply to the dual 3090s.
Parent - - By Lukas Cimiotti (Bronze) Date 2021-03-28 07:05

>Could it be a scaling issue or latency due to multiple sockets?


It's certaily no scaling issue - this is from a 2x Xeon 2679v4 reaching depth 44:


>This would also apply to the dual 3090s


The computer I used for comparison has dual 3080s. It should be 10 - 15% slower.
I guess he modified both engines to show higher node counts and recompiled them.
Parent - - By Carl Bicknell (*****) Date 2021-03-31 15:59 Upvotes 1
I can testify that amd 128 core's machine looks very fake.

I tested it in a way I knew he couldn't do anything about: on my private tactical testsuite with 100 test positions. I was using two machines:

1) Threadripper 3990X (64 cores, 128 threads)
2) Ryzen 3950X (16 cores, 32 threads)

His machines displayed far higher nps than either of mine, yet his machine had far inferior results on the test-suite and seemed to perform like a quad core.
Parent - - By Lukas Cimiotti (Bronze) Date 2021-04-01 07:41
Thank you.
A testsuite of course gives a better impression of playing strength. In fact it wouldn't be too difficult to falsify the indication of search depth.

Did you ever run your testsuite on lc0? I assume it might not be very good in tactics.
Parent - - By Carl Bicknell (*****) Date 2021-04-04 22:21 Upvotes 1
On my new suite, 100 positions, 3990X & 3090 Nvidia, 2 min per position

1. Black Diamond XI 99/100
2. Houdini 6 Tactical 84/100
3. Stockfish 13 82/100
4. Lc0 73/100
5. Fritz 15 and Fritz 16 58/100
Parent - - By Lukas Cimiotti (Bronze) Date 2021-04-05 05:57
Quite interesting.
I tried this: http://rybkaforum.net/cgi-bin/rybkaforum/topic_show.pl?pid=588335#pid588335
It looks to me like lc0 isn't as bad as I expected - and the network is really important.
Parent - By Carl Bicknell (*****) Date 2021-04-06 06:56
Yes, big nets do best.

When Lc0 gets a position right, it sees the solution very fast, like it just "knows" what move is best. But in other positions it's much slower than the alpha betas.
Parent - By MrKris (****) Date 2021-04-06 07:49

> It looks to me like lc0 isn't as bad as I expected - and the network is really important.


Sorry to say this, but I am suspecting your card.

Look at Nvidia.com's official designation for it!
https://www.nvidia.com/en-us/geforce/graphics-cards/30-series/rtx-3090/
Parent - - By MrKris (****) Date 2021-03-28 20:54
Its move 9 but this looks strange, maybe.

https://www.chess.com/computer-chess-championship#event=ccc-rapid-2021-finals
CPUs: 2 x AMD EPYC 7H12
[Event "CCC Rapid 2021: Finals (15|3)"]
[Site "Chess.com"]
[Date "2021.03.04"]
[Round "1"]
[White "Stockfish"]
[Black "Lc0"]
[Result "1/2-1/2"]
[ECO "A56"]
[GameDuration "00:31:48"]
[GameEndTime "2021-03-04T10:24:28.627 PST"]
[GameStartTime "2021-03-04T09:52:40.286 PST"]
[Opening "Benoni defense"]
[PlyCount "115"]
[Termination "adjudication"]
[TimeControl "900+3"]

1. d4 {book} Nf6 {book} 2. c4 {book} c5 {book} 3. d5 {book} g6 {book}
4. Nc3 {book} Bg7 {book} 5. e4 {book} O-O {book} 6. Bd3 {book} d6 {book}
7. h3 {book} e6 {book} 8. Nf3 {book} Nbd7 {book}
9. O-O {+0.75/43 94s, ev=0.75, d=43, pd=exd5, mt=00:01:33, tl=00:13:29, s=202151 kN/s,
n=18968263814, pv=O-O exd5... ... , tb=3553, R50=49, wv=0.75}
 
Parent - By The Wizard (***) Date 2021-03-31 21:47
Hi, I ran my 3990X on 64 cores so no HT ... You can better results tweaking the memory but if you are running the latest 4xPCI-e M2 drives it can drop there speed back to 3xPCI-e spec, well that is on my motherboard

Regards
Tony
Attachment: Capture.PNG (64k)
Parent - - By The Truth (**) Date 2021-04-12 03:02
I don't know why would someone even want to rent there, cpus are very overpriced. I can get a real 112c 224t epyc rome cpu on google cloud without having any doubts that the provider is screwing me for like less than 1 usd an hour on preemptible instance. Or if you want to get real crazy, azure offers 208c 416t Mv2 spot instances for like 5 usd an hour.

Also xeon v4s aren't really premium anymore, I can build a dual 2673 v4 40c 80t system for like less than $1k on chinese aliexpress. And you can even turbo mod some of these chinese motherboards so that cpus get higher clock speeds.
Parent - By Carl Bicknell (*****) Date 2021-04-13 17:43

> Or if you want to get real crazy, azure offers 208c 416t Mv2 spot instances for like 5 usd an hour.


How would that work if you had Aquarium on your local machine and wanted to use the 416T as an analysis engine?
- By Lukas Cimiotti (Bronze) Date 2021-05-31 11:27
Update:
amd 128 core is gone - but he returned calling himself "cluster king" and has fewer engines


Meanwhile he found out he misspelled his Xeon Paltinum / Platinum :twisted:
Also the CPU of his 2x 3090 liquid cooled Kingpin GPU computer changed - very strange.


I made a few more video clips of his engines - this is what he claims to be a 256 core cluster:
http://www.abdul-h.de/Clips/ckEpyc256.mp4
MN/s are much lower now, time to depth is better - but not better than a 64 core Threadripper 3990X

This is his 2x Xeon 8280:
http://www.abdul-h.de/Clips/ckXeon.mp4
MN/s are way too high for such a computer - it should be only slightly faster than immortal78's Xeon Platinum 8260M (.30 Ducats/min - s. screenshot above)

This is his 2x RTX 3090:
http://www.abdul-h.de/Clips/ck2x3090.mp4
KN/s are now much lower than before in my first post and time to depth is better. But it's worse than you would expect from such hardware.
This is my 2x 3080 which should be 10-20% slower:
http://www.abdul-h.de/Clips/Real_2x3080.mp4

This is what he claims to be a 4x RTX 2080ti computer:
http://www.abdul-h.de/Clips/ck4x2080ti.mp4
And this is one of my 2x RTX 2080ti computers:
http://www.abdul-h.de/Clips/Real_2x2080ti.mp4
This also makes clear the new NVidia generation of GPUs is not so much better than the previous.

In all lc0 tests the same network was used: LS16 =  PhoenixStein

Btw. the Chessbase cloud gets cpu and speed info of the pc you let for rent. But you can edit that info. If you do so, "Machine" is printed in grey, if you don't touch that info, it will be printed in black. There is one exception: If you choose "cluster controller" all info is black regardless what you write.

Has anyone tried some "cluster king" engines? Results from test sets or any other test would be highly apprecicated.
Up Topic The Rybka Lounge / Computer Chess / Fake computers in Chessbase Engine Cloud

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill