Author Topic: Will Corona take advantage of announced Xeon Phi accelerator ?  (Read 5093 times)

2012-11-13, 16:39:30

Dom74

  • Active Users
  • **
  • Posts: 92
    • View Profile
intel has announced the Xeon Phi accelerator :

http://pro.clubic.com/entreprises/intel/actualite-521991-intel-xeon-phi.html

Do you think Corona will/can take advantage of this hardware ?

2012-11-13, 18:31:28
Reply #1

Ondra

  • Administrator
  • Active Users
  • *****
  • Posts: 8896
  • Turning coffee to features since 2009
    • View Profile
Hi,
that depends on how much will it cost and how big speedup could be achieved. But it might be more viable than GPUs ;)
Rendering is magic.
Private scene uploader | How to get minidumps for crashed/frozen 3ds Max | Sorry for short replies, brief responses = more time to develop Corona ;)

2012-11-13, 19:03:22
Reply #2

Dom74

  • Active Users
  • **
  • Posts: 92
    • View Profile

Xeon Phi 3100 :
Disponibility first half of 2013,
nb. of core : unknow - 6 Go GDDR5 RAM - bandwith 240 Go/s - de 1000 Gigaflops in double precision - 225 watts  TDP - no price for now, but expected lower than 2000 $.

Xeon Phi 5110P :
Disponibility 28 january 2013,
60 x86 cores  - 8 Go GDDR5 RAM -  bandwith 320 Go/s - 1010 Gigaflops in double precision - 300 Watts  TDP - price 2649 $

Both can run X86 standard code, no need to re-write, so, eventually, compatible with all X86 softwares (?!)

2012-11-13, 19:32:36
Reply #3

Ondra

  • Administrator
  • Active Users
  • *****
  • Posts: 8896
  • Turning coffee to features since 2009
    • View Profile
6/8 GB memory is too little, 16/32 would be better :/. The price is ok-ish.

Gigaflops are largely unrelated measure of performance - if we were able to milk all gflops from current desktop CPUs, we could already render everything in real time.

The "can run x86, omgwtflol" is also more of a marketing advantage - you still need different binaries of your programs, because it does not have SSE (so you cannot directly runy ANY existing modern code with floating point operations). But binaries are produced by the compiler automatically, so I could in theory take corona and easily port it to PowerPC, ARM, smartphones, etc. What counts is the way you need to write the code to get acceptable speed. This is the reason GPU programming is hard. Even though you can write your programs with the same C++ syntax as when programming for CPU, you have to write it entirely differently. And I do not know how vast changes would be needed to milk the xeon phi ;).
Rendering is magic.
Private scene uploader | How to get minidumps for crashed/frozen 3ds Max | Sorry for short replies, brief responses = more time to develop Corona ;)

2012-11-13, 19:56:03
Reply #4

Dom74

  • Active Users
  • **
  • Posts: 92
    • View Profile
so, it was too good to be true...
Too sad.
Wait and see...

2012-11-13, 20:35:39
Reply #5

Ondra

  • Administrator
  • Active Users
  • *****
  • Posts: 8896
  • Turning coffee to features since 2009
    • View Profile
I'm not saying it is bad, it still looks more attractive, than GPU rendering ;)
Rendering is magic.
Private scene uploader | How to get minidumps for crashed/frozen 3ds Max | Sorry for short replies, brief responses = more time to develop Corona ;)

2012-11-13, 21:55:07
Reply #6

Dom74

  • Active Users
  • **
  • Posts: 92
    • View Profile
That wasn't so true before Corona.

2012-11-13, 21:56:11
Reply #7

Paul Jones

  • Active Users
  • **
  • Posts: 169
    • View Profile
8gb is plenty - well at least it does me, and my 2.5gb gfx card fits most of my interior scenes in with iray

2012-11-13, 22:39:59
Reply #8

Dom74

  • Active Users
  • **
  • Posts: 92
    • View Profile

2012-12-07, 01:55:20
Reply #9

Realish

  • Guest
Something to remember with both GPU rendering and this new phi processor is the RAM they have goes further.  This is because they don't need to store the OS + 3D program + whatever else on the GPU or Phi's RAM.  The only thing stored on the RAM on these will be the actual scene data with the rendering code needed.  This makes 8 GB on these cards work more like 12GB to 16GB or so on a full computer system.

2012-12-07, 12:17:13
Reply #10

maru

  • Corona Team
  • Active Users
  • ****
  • Posts: 9057
  • Marcin
    • View Profile
So what you're trying to say is that its effect will be the same as buying some more RAM for your pc? ;)

2012-12-07, 12:31:51
Reply #11

tomasd

  • Active Users
  • **
  • Posts: 19
    • View Profile
Something to remember with both GPU rendering and this new phi processor is the RAM they have goes further.  This is because they don't need to store the OS + 3D program + whatever else on the GPU or Phi's RAM.  The only thing stored on the RAM on these will be the actual scene data with the rendering code needed.  This makes 8 GB on these cards work more like 12GB to 16GB or so on a full computer system.

Given how much a 6GB (and I think that's the top model for now?) costs, wouldn't it be better to buy 1-2 new PCs instead?

2012-12-08, 05:34:47
Reply #12

Realish

  • Guest
Right now I know no programs that support the Phi, but once they do a single Phi might have the same speed as several top of the line PCs.  We will have to wait and see what programs take advantage of the new card and what kind of speed it gets in comparison to the top of the like CPU at the same time.  People will obviously go for the one that gives the most speed for their money.