Reply
 
LinkBack Thread Tools Display Modes
  #1   Report Post  
Old July 1st 14, 09:15 AM posted to sci.electronics.design,rec.games.chess.computer
external usenet poster
 
First recorded activity by ChessBanter: Jul 2006
Posts: 1,015
Default More CPU cores not always better

On 27/06/2014 11:47, John Devereux wrote:
Martin Brown writes:

On 26/06/2014 21:19, John Devereux wrote:


He might benefit from looking at the Scientific American article whose
infographic is already out of date (and cat's brain is missing LEDs)
but shows the raw compute power and storage capacity of various
computers and animals.


There is Moravecs classic essay from 1998, would like to see an update
of that.

http://www.transhumanist.com/volume1/moravec.htm

For the right problems the supercomputers have already won, but for
general pattern matching and inference the human brain still has the
edge by a long way. That is why Capcha still works (for now) and they
still farm out looking for galactic oddities to human volunteers.

http://www.scientificamerican.com/ar...ers-vs-brains/


Quick intro - this thread originated about the human mind and computer
simulations of it in s.e.d and how machine computational power is fast
approaching the point where desktop PCs rival human brain power.

Inspired by Moravecs 1998 essay I benchmarked my latest copy of Fritz14
on an i7-3770K ~$300 ~100k MIPS to add a new real point to the graphs.
That is about 330k MIPS/$1k or a bit less if it is whole system cost.

I have previously pointed out somewhat controversially in s.e.d that
more cores does not often translate to better performance on real
computing problems but in the case of 4 real cores with Intel
hyperthreading even I was surprised by the very pedestrian
"improvements" for chess as the extra cores were enabled.

CPU i7-3770K (stock speed 3.5GHz, 8GB ram, Win7 x64)

Fritz14 (x64) claimed ELO 3150 benchmarked

Cores Nodes/s

1 2.9M
2 5.6M
3 7.6M
4 9.6M

5 11.0M
6 12.0M
7 12.4M
8 11.1M (best out of three)

Cores 5-8 are hyperthreaded and this has never been good for chess.

Basically at around 6 cores active the external memory bus has saturated
and 8 cores is actually a worse performance than 7!

I didn't believe this 8 core result at first and so repeated it three
times giving results that ranged from 10.9M to the 11.1M shown.

The default settings are to use four cores and it is obvious why.
5 cores active is the best performance to power used tradeoff.

The results based on a 2012 launch and these benchmarks are that CPU
development is close to the most optimistic trend line on his graph.

I suspect but have not instrumented it that performance is hampered by
the local move generation causing writeback delays into main memory. It
might be helpful if some cache lines could be marked private for the
duration as internal working local memory for the cores.

I think the numbers for an i5-3570 would be broadly similar upto 4 cores
and the CPU price is considerably less. Guessing 500k MIPS/$1k
Anyone care to post actual tests?

I would be interested to see how a CPU design with 6 physical cores like
AMD's Thuban or the rare i7-980X (aka i9) performs on the same Fritz14
benchmark to determine how much of the problem is down to memory
bandwidth and how much to hyperthreading.

I'd also like to see how the new nominally 25% faster i7-4770 and the
eye wateringly expensive i7-3960X EE do on the Fritz14 benchmark as a
function of number of cores enabled.

Note for s.e.d now cross posted into rec.chess.computer

--
Regards,
Martin Brown
  #2   Report Post  
Old July 1st 14, 01:03 PM posted to sci.electronics.design,rec.games.chess.computer
external usenet poster
 
First recorded activity by ChessBanter: Jul 2014
Posts: 1
Default More CPU cores not always better

On Tue, 01 Jul 2014 09:15:04 +0100, Martin Brown
wrote:

On 27/06/2014 11:47, John Devereux wrote:
Martin Brown writes:

On 26/06/2014 21:19, John Devereux wrote:


He might benefit from looking at the Scientific American article whose
infographic is already out of date (and cat's brain is missing LEDs)
but shows the raw compute power and storage capacity of various
computers and animals.


There is Moravecs classic essay from 1998, would like to see an update
of that.

http://www.transhumanist.com/volume1/moravec.htm

For the right problems the supercomputers have already won, but for
general pattern matching and inference the human brain still has the
edge by a long way. That is why Capcha still works (for now) and they
still farm out looking for galactic oddities to human volunteers.

http://www.scientificamerican.com/ar...ers-vs-brains/


Quick intro - this thread originated about the human mind and computer
simulations of it in s.e.d and how machine computational power is fast
approaching the point where desktop PCs rival human brain power.

Inspired by Moravecs 1998 essay I benchmarked my latest copy of Fritz14
on an i7-3770K ~$300 ~100k MIPS to add a new real point to the graphs.
That is about 330k MIPS/$1k or a bit less if it is whole system cost.

I have previously pointed out somewhat controversially in s.e.d that
more cores does not often translate to better performance on real
computing problems but in the case of 4 real cores with Intel
hyperthreading even I was surprised by the very pedestrian
"improvements" for chess as the extra cores were enabled.

CPU i7-3770K (stock speed 3.5GHz, 8GB ram, Win7 x64)

Fritz14 (x64) claimed ELO 3150 benchmarked

Cores Nodes/s

1 2.9M
2 5.6M
3 7.6M
4 9.6M

5 11.0M
6 12.0M
7 12.4M
8 11.1M (best out of three)

Cores 5-8 are hyperthreaded and this has never been good for chess.

Basically at around 6 cores active the external memory bus has saturated
and 8 cores is actually a worse performance than 7!

I didn't believe this 8 core result at first and so repeated it three
times giving results that ranged from 10.9M to the 11.1M shown.

The default settings are to use four cores and it is obvious why.
5 cores active is the best performance to power used tradeoff.

The results based on a 2012 launch and these benchmarks are that CPU
development is close to the most optimistic trend line on his graph.

I suspect but have not instrumented it that performance is hampered by
the local move generation causing writeback delays into main memory. It
might be helpful if some cache lines could be marked private for the
duration as internal working local memory for the cores.

I think the numbers for an i5-3570 would be broadly similar upto 4 cores
and the CPU price is considerably less. Guessing 500k MIPS/$1k
Anyone care to post actual tests?

I would be interested to see how a CPU design with 6 physical cores like
AMD's Thuban or the rare i7-980X (aka i9) performs on the same Fritz14
benchmark to determine how much of the problem is down to memory
bandwidth and how much to hyperthreading.

I'd also like to see how the new nominally 25% faster i7-4770 and the
eye wateringly expensive i7-3960X EE do on the Fritz14 benchmark as a
function of number of cores enabled.

Note for s.e.d now cross posted into rec.chess.computer


Hyper threading is not the same as a core. It is typicaly turned off
on SQL servers because it degrades performance.


Cheers
  #3   Report Post  
Old July 2nd 14, 11:22 AM posted to sci.electronics.design,rec.games.chess.computer
external usenet poster
 
First recorded activity by ChessBanter: Jul 2006
Posts: 1,015
Default More CPU cores not always better

On 01/07/2014 13:03, Martin Riddle wrote:
On Tue, 01 Jul 2014 09:15:04 +0100, Martin Brown
wrote:

On 27/06/2014 11:47, John Devereux wrote:
Martin Brown writes:

On 26/06/2014 21:19, John Devereux wrote:


He might benefit from looking at the Scientific American article whose
infographic is already out of date (and cat's brain is missing LEDs)
but shows the raw compute power and storage capacity of various
computers and animals.

There is Moravecs classic essay from 1998, would like to see an update
of that.

http://www.transhumanist.com/volume1/moravec.htm

For the right problems the supercomputers have already won, but for
general pattern matching and inference the human brain still has the
edge by a long way. That is why Capcha still works (for now) and they
still farm out looking for galactic oddities to human volunteers.

http://www.scientificamerican.com/ar...ers-vs-brains/


Quick intro - this thread originated about the human mind and computer
simulations of it in s.e.d and how machine computational power is fast
approaching the point where desktop PCs rival human brain power.

Inspired by Moravecs 1998 essay I benchmarked my latest copy of Fritz14
on an i7-3770K ~$300 ~100k MIPS to add a new real point to the graphs.
That is about 330k MIPS/$1k or a bit less if it is whole system cost.

I have previously pointed out somewhat controversially in s.e.d that
more cores does not often translate to better performance on real
computing problems but in the case of 4 real cores with Intel
hyperthreading even I was surprised by the very pedestrian
"improvements" for chess as the extra cores were enabled.

CPU i7-3770K (stock speed 3.5GHz, 8GB ram, Win7 x64)

Fritz14 (x64) claimed ELO 3150 benchmarked

Cores Nodes/s

1 2.9M
2 5.6M
3 7.6M
4 9.6M

5 11.0M
6 12.0M
7 12.4M
8 11.1M (best out of three)

Cores 5-8 are hyperthreaded and this has never been good for chess.

Basically at around 6 cores active the external memory bus has saturated
and 8 cores is actually a worse performance than 7!


[snip]

I think the numbers for an i5-3570 would be broadly similar upto 4 cores
and the CPU price is considerably less. Guessing 500k MIPS/$1k
Anyone care to post actual tests?

I would be interested to see how a CPU design with 6 physical cores like
AMD's Thuban or the rare i7-980X (aka i9) performs on the same Fritz14
benchmark to determine how much of the problem is down to memory
bandwidth and how much to hyperthreading.

I'd also like to see how the new nominally 25% faster i7-4770 and the
eye wateringly expensive i7-3960X EE do on the Fritz14 benchmark as a
function of number of cores enabled.

Note for s.e.d now cross posted into rec.chess.computer


Hyper threading is not the same as a core. It is typicaly turned off
on SQL servers because it degrades performance.


I am well aware that hyperthreading is inferior and not the same as a
physical core which is why I would like to see the figures for the
handful of hex physical core devices on the same benchmark.

The other slightly curious thing is that adding cores 3 and 4 contribute
the same performance increase to within measurement error.

However, it seems likely that forcing maximum logical CPUs to be 1 or 2
more than the physical core count still has some benefit for chess.

The same might also be true in SQL server clusters. Mickeysoft are
rather coy about the right setting for MAXDOP in their documents.

Hyperthreading is a marketing man's dream for specsmanship. It is a
shame that cache invalidation often wipes out the performance gain.

--
Regards,
Martin Brown
Reply
Thread Tools
Display Modes

Posting Rules

Smilies are On
[IMG] code is On
HTML code is Off
Trackbacks are On
Pingbacks are On
Refbacks are On


Similar Threads
Thread Thread Starter Forum Replies Last Post
Deep Fritz for Core2 Duo? tonysin rec.games.chess.computer (Computer Chess) 6 June 2nd 09 04:50 PM
Stupidest game of the month. 11 move Mate. Sanny rec.games.chess.computer (Computer Chess) 49 July 18th 08 12:52 AM
What is Rating of Rybka and other Programs? Sanny rec.games.chess.computer (Computer Chess) 65 June 22nd 08 06:06 AM
What is Rating of Rybka and other Programs? Sanny rec.games.chess.misc (Chess General) 63 June 20th 08 04:55 AM
What is Rating of Rybka and other Programs? Sanny alt.chess (Alternative Chess Group) 42 June 19th 08 03:47 AM


All times are GMT +1. The time now is 09:47 AM.

Powered by vBulletin® Copyright ©2000 - 2019, Jelsoft Enterprises Ltd.
Copyright 2004-2019 ChessBanter.
The comments are property of their posters.
 

About Us

"It's about Chess"

 

Copyright © 2017