newton on gpu or discrete HPC

A place to discuss everything related to Newton Dynamics.

Moderators: Sascha Willems, walaber

Re: newton on gpu or discrete HPC

Postby Julio Jerez » Sat Sep 29, 2018 7:51 am

not is not accuracy is the number of operation follows different order because becaus ethe contact array come in different order each time a for a multithreaded than is in single threaded.

this has bothered me a lot and I tried to make it deterministic means losing al most all teh performance because afte each threaded operation it would nee to sort the result lexicographically.
this is not worth the effort.

It is still a mystery why I get to go to rest at 16 and some time even at 12, an I am using the same sse solver, but you need 20.
but at least when is does goes to rest and some point.
and that allowed us to have a fix point to measure the performance of other solvers.

and as you said this really is an stress test of accuracy rather than of performance which si my priumary objective with the solver. under no circunatce I will accept diffrent result with GPU than with CPU given the same engine settings.
Julio Jerez
Moderator
Moderator
 
Posts: 12249
Joined: Sun Sep 14, 2003 2:18 pm
Location: Los Angeles

Re: newton on gpu or discrete HPC

Postby JoeJ » Sat Oct 06, 2018 4:40 pm

Seems this is the first game requiring AVX: https://www.dsogaming.com/news/assassins-creed-odyssey-will-not-work-on-cpus-that-do-not-support-avx/#more-118161
Although i guess they'll do a patch and this is just lack of conversation.
User avatar
JoeJ
 
Posts: 1453
Joined: Tue Dec 21, 2010 6:18 pm

Re: newton on gpu or discrete HPC

Postby Julio Jerez » Sat Oct 06, 2018 5:10 pm

I like that they took that bold move.
It will put more pressure on Intel and And to make avx actually more efficient. specially Amd which does not really suport it at the microcode level.
in many cases avx code on AMD cpus actually run slower than plain vanilla sse and on Intel is actually marginally faster than SSE.
To get actual speed up you need to have very large arrays of vector that do the same thing and that requires sofisticated transformations that usually cost more than the gains.

It took more than ten years for see to be better than x87, with the core duo and it seems avx is taking even longer. I do not see avx512 been adopted anytime soon with and not adopting it.

I welcome the move, even thought I am pretty sure they are running significantly slower on consoles.
Julio Jerez
Moderator
Moderator
 
Posts: 12249
Joined: Sun Sep 14, 2003 2:18 pm
Location: Los Angeles

Re: newton on gpu or discrete HPC

Postby JoeJ » Sat Oct 06, 2018 5:22 pm

Hey, now i have a PC that is still fast but can't run new software because it's too old.
Feels like a Macintosh now :D
User avatar
JoeJ
 
Posts: 1453
Joined: Tue Dec 21, 2010 6:18 pm

Previous

Return to General Discussion

Who is online

Users browsing this forum: No registered users and 15 guests

cron