Welcome, Guest. Please login or register.

Author Topic: AmigaOne X1000  (Read 8011 times)

Description:

0 Members and 1 Guest are viewing this topic.

Offline DAX

  • Full Member
  • ***
  • Join Date: Jun 2010
  • Posts: 163
    • Show all replies
Re: AmigaOne X1000
« on: June 26, 2010, 10:43:07 AM »
Quote
no fpu...can't beat 1 GFlops

As we are entering the era of heterogeneous computing any additional data crunching facility the system has acces to is good.

Talking about 3D rendering (my field) and GP-GPU, there have been a lot of misconceptions flying around, with fancy RT demonstrations that quite frankly demonstred very little. The problem there, is that they use a lot pre-baked data while GP-GPU finishes the rest (AtiCinema2.0 is a good example), but when the rendering is 100% handled without baked data (as shown in some horrid raytracing examples), the quality of the rendering drops dramatically and can't hold a candle to non RT algorithms, while when the CPU intervenes in a hybrid mode, the slow downs are as heavy as it's the improvement in picture quality you set (one example would be the recent "mostly gpu based" Quicksilver engine for 3DS max, VS the same scene done in plain mental-ray, but it gets deeper than that).

To solve this dilemma ex Nvidia/Ati employees have formed Caustics Graphics, a company that produces a board which only intervene into a more complex process in order to aid where CPUs and GPUs are not very good at, preparing data for their consumption (heterogeneous computing at its finest we might say, everything is used) more infos about this here:

http://www.vizworld.com/2009/09/caustics_3dsmax_demonstration/ (whatch the video and listen to what they ave to say on the subject)

Considering only some data is sent to the board for prepping (the rest is done by the CPU and GPU) the bandwidth required is modest and I wonder if a minicluster of Xmos chips could handle the same kind of data (that card is expensive, you can build a 256 Xmos cluster for less) and work together with both CPU and GPU in a similar manner (or at least have the trio achieve a 1 GI frame per second result, which is still very good).

On Xmos forum you can read about the SoftFloat they implemented, it is extremely fast for being "soft", and on a cluster of 500Mhz chip you might still get what you need for "trio" operartions as the one above (or other, a different thing would be if it had to handle everything but this is not the case).

@Pyromania
In any case go for it, it's gonna be a fun exercise and you will grant the original creator's wish :)
« Last Edit: June 26, 2010, 11:07:16 AM by DAX »