OpenCL wishlist :)

  • 9 Replies
  • 213 Views

0 Members and 1 Guest are viewing this topic.

Offline R-TEAM

  • *
  • Fractal Friend
  • **
  • Posts: 12
« on: July 10, 2020, 05:12:18 PM »
Hi,
as i have ATM more free time - i play around with the new Mandelbulber ...
Use my RadeonPro GPU - still takes 18h to render .... would like to use my other workstations, but miss i something or is NetRender atm only for CPU only .. ?
And it is possible to Pause/UnPause the rendering ?
And the "all time wish" for users with an "pro" GPU .... double precision OpenCL rendering ;)

Regards

Linkback: https://fractalforums.org/mandelbulber/14/opencl-wishlist/3656/

Offline buddhi

  • *
  • Fractal Feline
  • **
  • Posts: 157
    • Mandelbulber GitHub repository
« Reply #1 on: July 10, 2020, 07:57:36 PM »
NetRender for OpenCL actually works only with animations. For still frames of course would be nice to implement. But it is not easy.
About double precision, I don't have GPU capable for this stuff. GTX 1060 is way to slow and limited to run big kernels with double precision number. This is one of the reasons why I haven't started development of this. Second thing, even very expensive GPUs are still slower with double precision than good CPUs.

Offline buddhi

  • *
  • Fractal Feline
  • **
  • Posts: 157
    • Mandelbulber GitHub repository
« Reply #2 on: July 10, 2020, 08:04:09 PM »
After reading this I have changed my mind about speed: https://www.tomshardware.com/news/AMD-announced-Radeon-Pr
But price for cards is way toooooooo high.

Offline R-TEAM

  • *
  • Fractal Friend
  • **
  • Posts: 12
« Reply #3 on: July 10, 2020, 10:22:06 PM »
about speed imho a wrong argument ... (to slow)
My card (as all highend pro cards) use in double precision only half the speed as in single precision - thats a slowdown ...
But even my 12 Core Xeon cpu (an older model) cant match with half the speed ... and i dont count the price for an 28 or higher core count CPU ...
So imho double precision on card where use this full (half speed as single precision) should always be faster and more economical than CPU rendering
Consumer cards are horrible at anything more that single precision - but this a different story .... (the cards simply maked for gaming - not for working with it... and the working capability you pay on the pro cards [radeonpro/quadro] )

Offline R-TEAM

  • *
  • Fractal Friend
  • **
  • Posts: 12
« Reply #4 on: July 11, 2020, 12:31:24 AM »
and to get "real" numbers :
Brand NEW Ryzen Threadripper 3990X DP > 3.2 TFlop  at ~4000€ + proffessional Mobo + RAM (if you would not use it as a toy...)
my 6 ! Year Old AMD RadeonPro w8100 DP > 2.11 TFlop at ~480€ (Amazon) [and an free PCIe 3.0 16x Slot]

So it is not "new" that GPU rendering is much faster ..... only if you compare CPU Vs cheap pro or consumer cards where was never designed for this
Doing math on the edge of possibility with toys ...........

Offline buddhi

  • *
  • Fractal Feline
  • **
  • Posts: 157
    • Mandelbulber GitHub repository
« Reply #5 on: July 11, 2020, 11:02:01 PM »
...but it doesn't change fact, that I have no budget for any of these pro cards. New one costs as much as my whole workstation.

Offline R-TEAM

  • *
  • Fractal Friend
  • **
  • Posts: 12
« Reply #6 on: July 11, 2020, 11:50:13 PM »
this ... is a different point ;) (and an valid)
but "to slow" is not an valid point :)

and yes ..... i am not happy with the price as well ....
But this dont change the price magical :P
[planing my next workstation with ~15k€ ... so an pro card for 2~3k€ is not an so big part ..... but i am still not sure Intel Xeon or AMD ... and the actual high-end AMD pro card is more mhee... so it will still take time..]

> and i still dont see an strong point of not supporting DoublePrecision - only you cant use it sensefull  on your card (as it have to low DP Flops) m this dont mean you cant develop it , it works even on cheap cards - only slow - this will change maybe over time (Nvidia have last time changed the bahavor of true 10bit output from Pro cards only to even selected consumer cards .. so nothing is set in Stone) - and other can use DP to make more fantastic images

Offline mclarekin

  • *
  • Fractal Freak
  • **
  • Posts: 784
« Reply #7 on: July 12, 2020, 03:08:59 AM »
i think the main point is that most users do not have high end cards ATM.  Double precision openCL is on the ever growing TODO list, but without extra programmers joining the project, we  mainly work on what will benefit most current users in the available time we have for this hobby.

Offline R-TEAM

  • *
  • Fractal Friend
  • **
  • Posts: 12
« Reply #8 on: July 13, 2020, 12:34:19 PM »
i dont say a majority of Mandelbulber users have an card with an high SP/DP OpenCL ratio .... but this point you can say to many additions, that only a small amount of users benefits from new feature A or B ...
anyway - it is a free project so the devs have to decide where the way is going ...

Another suggestion > OpenCL with GPU AND CPU , as actual more users would have an multicore CPU with enough power to benefits from this (even if the CPU only use 20% from the GPU OpenCL power ...)

Offline R-TEAM

  • *
  • Fractal Friend
  • **
  • Posts: 12
« Reply #9 on: July 13, 2020, 03:57:23 PM »
Oh - and this musst -> a BIG THANKS to all developers where make this wounderfull programm happen :)
It works realy stable and even in high stress situations ... have my best rendered pictures on ArtStation ( https://www.artstation.com/artwork/9O64N ) and the last Picture i have rendered with MonteCarlo in ~16h ...
And have in this time show YouTube , browse the web - open/close different browsers, chat .... all have worked nice (o.k. - little shutter now and then on window opening or similar - but understandable) and stable - no crash or data lost.

Regards


xx
Mandelbulber v2 2.12 - OpenCL

Started by buddhi on Downloads

4 Replies
2076 Views
Last post December 05, 2017, 09:52:27 PM
by knighty
clip
OpenCL only pink in 2.12/2.13

Started by Micha1982 on Mandelbulber

1 Replies
271 Views
Last post April 01, 2018, 02:18:34 AM
by Micha1982
thumbup
Mandelbulber 2.2.16 OpenCL

Started by piotrv on Mandelbulber

2 Replies
610 Views
Last post December 25, 2018, 01:09:32 PM
by Caleidoscope
xx
OpenCL Netrender: GPU/CPU

Started by greytery on Mandelbulber

6 Replies
689 Views
Last post June 08, 2019, 04:16:40 AM
by Tas_mania
xx
MC DOF OpenCL Test

Started by buddhi on Fractal Image Gallery

0 Replies
334 Views
Last post September 04, 2017, 09:28:21 PM
by buddhi