Not logged inRybka Chess Community Forum
Up Topic The Rybka Lounge / Computer Chess / Booted at 5Ghz
- - By The Wizard (***) Date 2019-04-06 15:14
Hi, With the new Omega motherboard and leaving certain voltages set on AUTO for the mobo to decide I managed to get the i7980XE to boot at 5Ghz on my water cooling setup something I could not do with the previous board. Don`t get me wrong it is not stable enough for a chess engine running flat out but I managed to validate easy enough.
       The board decided itself to put 1.4V into the CPU :eek::evil::twisted: but to my surprise it only reached 80C  .... It seems I might better overclocks if I drop the CPU cache speed down a lot. TBH though I doubt I will ever get it chess engine stable at 5Ghz even when I cool it sub zero but never say never

I changed my main drive to a Samsung 970 EVO PLUS and have all my chess stuff now on the Samsung 960 PRO ... I still have an old PCi OCZ Revo 3x2 and a OCZ Vertex SSD connected for backups etc

https://valid.x86.fr/8z3wlf
Parent - - By Labyrinth (*****) Date 2019-04-07 02:14
Are you trying to reach 90 mN/sec? :-)
Parent - - By The Wizard (***) Date 2019-04-09 23:42
Hi Labyrinth, That is my intention and I suspect it is not all about CPU Mhz .... A fresh install of windows and basic drivers I think would crack it. Maybe running the benchmark in safe mode might do it?

Regards
Parent - By Labyrinth (*****) Date 2019-04-10 08:20
You could try setting the priority of the engine exe higher, see if that helps. In my experience 'high' and 'realtime' tend to cause massive system instability though. I think the last time I tried realtime on a program I got a bluescreen in about 5 seconds.
Parent - - By Peter Grayson (****) Date 2019-04-08 10:10 Upvotes 1
It would be interesting to measure in reality how much faster a chess engine is in solving a position for each increment of CPU speed for example each 0.2 GHz and if possible to also measure wattage although I would start with comparing the stock speed to peak speed you intend to run. Core-Temp does indicate power on my i7-8700 and dual Xeon machines that seems to accurately reflect the CPU power consumption.

By running a single position solving test say 100 times and averaging the time to solve would indicate the engine's degree of variability when run with more than one thread together with average solving time. It may highlight the value of overclocking could be lost to a certain degree when there is no certainty the engine would find the correct move much faster for any given run. kN/s as a measure can look impressive but does not truly reflect an engine's performance.

Peter
Parent - - By The Wizard (***) Date 2019-04-10 00:02
Hi Peter, I use HWiNFO from https://www.hwinfo.com/ but run the 64bit portable version over at https://www.fosshub.com/HWiNFO.html

I will try to find a few positions and run the tests you ask ..... ATM I am try to see what is the lowest CPU voltage I can run an overclock at. With hyperthreading on CPU temps are approx 8c higher with 1.25V .... Ambient temps are not friendly and will only get worse as the summer approaches

Regards
Parent - - By Peter Grayson (****) Date 2019-04-10 16:43

> Ambient temps are not friendly and will only get worse as the summer approaches


That was the issue I found running a dual Xeon machine in a thermally efficient house! Windows open - heat out - bugs in! Late last year I settled on a 65 watt i7-8700 when flat out its heat contribution seems minimal but of course the time to solve is down but not using it for any competitive chess so no problem. Just 65 watts versus 300 watts per hour of the dual Xeon for the CPUs when the economics of it for 24 hour running doesn't make sense. Time is the issue but it does come at a cost.

I've attached my file on the measurements and the database with the position x100 I used with Stockfish 9 run in Chessbase GUI 14 64 bit. Syzygy bases need to be disabled and I used just 1 Gb hash because was more interested in the relative CPU performance. It was very enlightening!

Thanks for the links. I'll have a closer look at that software. Looks interesting.

Peter
Attachment: ReidP2SF9HTTest.zip - The files (35k)
Parent - - By The Wizard (***) Date 2019-04-11 06:08
Hi Peter, I ran a quick test of the first 25 positions basically to make sure I got it right ..... Watching the engine got me a bit confused seeing just how different HT on or off can be and not always in the direction I would have expected ... I will get around to doing it at different speeds as I need to turn TURBO off so that 26x100 is really 2.6Ghz .....  Also unless you force it to (as I do) instead of all cores running the same speed and voltage it can set core speeds and voltages differently  ... There is also more options in the BIOS that can be altered and may make a difference

Regards
Parent - By Peter Grayson (****) Date 2019-04-11 09:52
Thanks Tony?(not sure how good my memory is!)

I suspect the results reflect what I found. The 29% kN/s gain obtained on the dual Xeon machine looked about right for the 30% potential gain of HT-on to HT-off but for chess engines the important measurement is time to solve that on my machines, including the older quads, does not reflect the kN/s difference but instead a much narrower difference, 8.6% for the dual Xeon, confirming the time gain is not reflective of the kN/s difference that tends to flatter to deceive.

The inherent multi-thread variability of the engine is the culprit of the solve time variations and the averaged narrower speed by time measurement differential reflects the overhead cost of hyper-threading. The multi-thread variability is why the test must be run so many times when the more runs gives a more accurate measure of the dfference.

It will be interesting though to see the final results achieved with CPU performance optimised.

Peter
Up Topic The Rybka Lounge / Computer Chess / Booted at 5Ghz

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill