[wplug] video card power consumption -- performance comparison ...

[Bryan J. Smith]

On Mon, 2007-09-10 at 10:40 -0400, Bryan J. Smith wrote:
> Intel's open source driver's support is not only pathetic, but Intel
> itself _hordes_ IP.  nVidia, for the longest time, could _not_ release
> the AGPgart code to the kernel because it contained Intel IP.  That's
> why only nVidia's closed source driver game with the AGPgart for its
> nForce chipsets.  Once PCIe arrived, Intel stopped treating AGP like a
> "trade secret," and the AGPgart for nForce chipsets went into the
> kernel.  Intel does _not_ provide a kernel driver, which kills
> performance.  Beyond that, they have _failed_ to support various
> features in their i945/955 GPUs.

BTW, when I say "kernel driver," I mean the "kernel-memory driver."
There are hooks and comments in the Intel open source driver for this.
It confuses most people.

Anyone who has looked at the Windows driver knows the difference.  ;)

In a nutshell, when nVidia got their cease'n desist letters, they
basically said, "oh, to hell with the 3rd party IP issues."  They then
went off and created their "unified object code" and wrappers so they
worked on any OS.  This included the necessary kernel-memory driver, of
which has a lot of Intel IP.

I have been a _huge_ fan of AMD buying ATI (even if it wasn't fiscally
ideal), and AMD _had_ to buy ATI.  The reason?  AMD offers a _true_,
_hardware_ I/O MMU and memory manager for "expansion cards that try to
act like CPUs" -- such as a GPU.

AMD can and will make a HTX video card shortly, starting with mainboard
integrated.  Once the GPU is on the HyperTransport interconnect, it
_removes_ all of those _stupid_ "software hacks" that Intel forces
everyone to use (because they continue to _lack_ a real "system
interconnect" with cache/memory coherency between devices).

Until then, Intel's software IP infects everything that is "high
performing."  Because without those hacks, you don't get squat
performance-wise.  Why?  Because the CPU and GPU have coherency issues,
things that require software hacks, things that Intel considers "trade
secrets."

Software hacks AMD CPUs don't need _if_ the GPU was designed as a "peer"
node on the system interconnect.  Something Intel can't currently do.

That's the #1 issue with the kernel driver right now, and the reason why
it's not GPL.  And it's Intel's fault, not ATI or nVidia's.  Intel
avoids putting that IP in the drivers, and performance is horrendous.
Matrox doesn't bother either.  With all the IP isues, I don't blame
them.

Heck, Microsoft has sucked up a lot of OpenGL-related IP, so _any_ open
source OpenGL implementation has indemnification issues as it is.  With
a legal, licensed, closed source implementation, that solves that
problem as well.

The only way around that is to get away from OpenGL.  But even then,
most of the IP on OpenGL would still apply to that non-OpenGL
implementation.

With is why most of this "open source" non-sense is rather misguided.
Especially when Intel is praised (while being part of the problem ;).
  
-- Bryan

P.S.  Yes, I have a chip on my shoulder from dealing with this on Linux
for 10 years.  Especially when someone at a client starts in with, "why
don't we use Intel and the open source drivers?"  Typically the "because
the application won't run" is good enough, although then they try to say
it's because of "proprietary functions" until I start throwing the
OpenGL 1.3, 2.0 and officially spec'd ARB extension at them, and the
massive _glut_ of what Intel does not do.  ;)



-- 
Bryan J. Smith         Professional, Technical Annoyance
mailto:b.j.smith at ieee.org   http://thebs413.blogspot.com
--------------------------------------------------------
        Fission Power:  An Inconvenient Solution