[extropy-chat] Using Graphic controllers

[Eugen Leitl]

On Sun, Apr 18, 2004 at 12:28:29PM +1200, Paul Bridger wrote:
> Dan Clemmensen wrote:
> 
> >Another interesting point: From the point of view of raw computational 
> >power, the sum of all the 3D graphics cards in the world probably 
> >exceeds the sum of all CPUs in the world. Sure, the graphics cards are 

It depends how you measure performance. I presume everybody is aware of
http://www.gpgpu.org/

GPUs have a usually rather impressive memory bandwidth and good parallelism
(broad buses). On the minus side, they're still a von Neumann design (in
general terms), and the memory is expensive, hence limited to small
increments. They've peaked out on power density and power supply (demand
ramps up higher than power supply can provide, even capacitor-buffered).

> >special-purpose, but what if you had the resources to carefully create 
> >algorithms that are well suited to graphics cards? basically, this 

Sure, so you get one order of magnitude more out of it. On a single
architecture, the deployed hardware base is heterogenous, in flux, and
there's no common API. 

In terms of ROI, it makes more sense to focus on CPUs, and stuff which
streams through memory, keeping hot spot barely within cache. You cluster
above using short-distance signalling mesh and WAN interconnects.

> >requires that you map your problem onto the space that is easily 
> >addressed by graphics cards. The extremely obvious problem is graphics 
> >rendering (DUH!) This has actually been done: you can use graphics 
> >cards to run the POV-RAY backend. When you do this, a few (10?) 
> >machines with graphics cards can render scenes that would otherwise 
> >take hundreds of high-end CPUs.
> 
> The major problem with the unused GPU (Graphics Processing Unit) power 
> is that most of it is primitive fixed-function hardware. Newer hardware 
> is increasingly general purpose (I'm talking about vertex/pixel 
> programs). My guess is that the number of problems that can be mapped 
> into the space of fixed-function hardware is fairly limited. However, 
> programmable hardware raises incredible new possibilities. The only 
> thing holding back photorealistic rendering in realtime today is 
> programmer skill.
> A large restriction on the computation you farm off to your unused GPU 
> resource is that it must be able to be computed in parallel. This is the 
> reason that modern graphics cards have higher complexity than modern 
> CPUS -- graphics processing is an inherently parallel process and the 
> structure of the cards mirrors this. Modern graphics cards run >500MHz 
> (but have high parallelism) whereas modern CPUs are upwards of 3GHz.

-- 
Eugen* Leitl <a href="http://leitl.org">leitl</a>
______________________________________________________________
ICBM: 48.07078, 11.61144            http://www.leitl.org
8B29F6BE: 099D 78BA 2FD3 B014 B08A  7779 75B0 2443 8B29 F6BE
http://moleculardevices.org         http://nanomachines.net
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 198 bytes
Desc: not available
URL: <http://lists.extropy.org/pipermail/extropy-chat/attachments/20040418/b3868af8/attachment.bin>