Re: OT: nVidia driver [was: Wish list] -- understanding the GPU market

[Bryan Smith]

Matrox has been proprietary.
ATI is now proprietary.
The XFree86 4.0+ model introduced a loadable module option, which
allowed binary drivers.

nVidia just adds some kernel memory and AGPgart interfaces.
Until recently, those were Intel "trade secrets" nVidia could not
disclose.

As far as the unified driver, Matrox, nVidia and now ATI have the
same issue, Intel, Microsoft and a host of other companies' lawyers
would have a field day if they were GPL'd.

ATI should be commended for attempting to make a "clean room" DRI/
GLX implementation.  But eventually they had to give in, and started
withholding specifications as of R300 (Radeon 9500+).  And even before,
many 2D and 3D interfaces were _never_ published by ATI.

What really gets to me here is the _real_market_conditions_ involved.

We're not talking about a product that is released and modern for 2-5
years.  We're talking about a product that is _obsolete_ in less than
_12_months_!

GPUs double in speed 2-3 times faster than CPUs.  The drivers are
under development alongside engineering under NDA, and that's not
going to change.  If they opened up the driver model, then they'd
still either need to A) have you sign a NDA, or B) you'll see
drivers for Linux come out some 6+ months _after_ release.

I think there are lessons to be learned from ATI's prior attempts.

In reality, we're still talking an "open standard" in GLX, with
ARB extensions and in a few cases, yes, some GPU-specific extensions
(that get rolled into ARB much faster than DirectX where many
_never_ become part of the spec).

The concept that leading-edge video drivers will _ever_ be GPL is 
very slim-to-none.  And as far as the license, the drivers are
unified, and written for _all_ platforms, so there is the argument
that they do not require Linux (and Linus' comments have been used
similarly for justifying binary-only WLAN objects).

In reality, yes, I'd like to see nVidia's memory interfaces in the
kernel GPL'd.  And now that the _standard_ PCI-Express is here from
the PCI Standards Group (AGP was _never_ a standard, but an Intel
trade secret of PCI), I hope Intel takes off some chains on nVidia.

But there is just too much IP involved, and too many NDAs.  GPUs
are far, far, _far_ more complex than even CPUs these days -- and
the IP in the software even more so and closely tied.

There is some open source GLX code out there for both ATI and nVidia
-- but you're _never_ going to get the "cutting edge" under the model.


-- 
Bryan J. Smith                                     b.j.smith(a)ieee.org 
--------------------------------------------------------------------- 
It is mathematically impossible for someone who makes more than you
to be anything but richer than you.  Any tax rate that penalizes them
will also penalize you similarly (to those below you, and then below
them).  Linear algebra, let alone differential calculus or even ele-
mentary concepts of limits, is mutually exclusive with US journalism.
So forget even attempting to explain how tax cuts work.  ;->

OT: nVidia driver [was: Wish list] -- understanding the GPU market

[Dave Jones]

On Tue, Jun 07, 2005 at 11:43:36PM -0500, Bryan J. Smith wrote:
>  nVidia just adds some kernel memory and AGPgart interfaces. 
Plus god knows what else. Take a look at the size of the binary
part of their driver. It's (to the best of my knowledge) the
largest kernel module available for Linux.

>  Until recently, those were Intel "trade secrets" nVidia
> could not
>  disclose. 
I'm not sure where you're getting this.

>  ATI should be commended for attempting to make a "clean
> room" DRI/
>  GLX implementation.  But eventually they had to give in 
The r200 work was actually done by precision insight, under
sponsership from The Weather Channel, not ATI.

>  The concept that leading-edge video drivers will _ever_ be GPL is 
>  very slim-to-none. 
The folks reverse engineering the r300 cores have come along in leaps
and bounds. From what I hear it's good enough to play quake and such
already.  Having to reverse engineer the hardware does mean we're
at best going to be one generation of card behind though, but the
picture isn't totally bleak..

>  In reality, yes, I'd like to see nVidia's memory interfaces
> in the
>  kernel GPL'd.  And now that the _standard_ PCI-Express is here from
>  the PCI Standards Group (AGP was _never_ a standard, but an Intel
>  trade secret of PCI) 
If you mean this in the sense "Was never a standard that PCI-SIG made
people pay for to see", you are correct. But there was nothing secret
about it at all. The fact that it was freely downloadable was one of the
reasons that Linux ended up with agpgart support so quickly.

		Dave

OT: nVidia driver [was: Wish list] -- understanding the GPU market

[Nicolas Mailhot]

On Mer 8 juin 2005 6:43, Bryan J. Smith a écrit :

>  We're not talking about a product that is released and modern for
> 2-5
>  years.  We're talking about a product that is _obsolete_ in less than
>  _12_months_! 
This is a very misleading statement.
The core driver architecture has a much longer lifetime.
Ati/NVidia are not adding a thin closed layer above a common core. They're
systematicaly bastardising this common core to avoid having to distingish
between the secret and public parts. In other words, they're lazy.

Graphic cards are not the only parts in a computer with a short lifetime.
Srangely other members of the industry have little problem standardising
the hardware interface.

Regards,

-- 
Nicolas Mailhot

OT: nVidia driver [was: Wish list] -- understanding the GPU market

[Alan Cox]

On Tue, Jun 07, 2005 at 11:43:36PM -0500, Bryan J. Smith wrote:
>  ATI should be commended for attempting to make a "clean
> room" DRI/
>  GLX implementation.  But eventually they had to give in, and started
>  withholding specifications as of R300 (Radeon 9500+).  And even before,
>  many 2D and 3D interfaces were _never_ published by ATI. 
Actually its not quite as simple as some people may think here. ATI maintain
a lot of stuff and help with many things. The lack of an ATI open R300 driver
is at least partly certain Linux organisations fault for not taking up 
opportunities rather than ATI's. (and more I cannot say in public)

>  The concept that leading-edge video drivers will _ever_ be GPL is 
>  very slim-to-none.  And as far as the license, the drivers are 
Its more an inevitability than the reverse. Commoditisation means that very
soon all the 'must keep secret' IP will be only relevant to real time
ray tracing. What happens to the Radeon 9800 when Intel 9xx/VIA/etc graphics
on chip are good enough for gamers. Leading-edge video will be for VR nuts.

(And guess why Nvidia and ATI are in the chipset business nowdays)

>  the PCI Standards Group (AGP was _never_ a standard, but an Intel
>  trade secret of PCI), I hope Intel takes off some chains on nVidia. 
Hardly. AGP is a published open specification. People like nVidia do some
really clever tricks with it for performance like the DMA contexts but nothing
major we don't actually understand AFAIK.

>  There is some open source GLX code out there for both ATI and nVidia
>  -- but you're _never_ going to get the "cutting edge" under the model.
Which would be why folks are currently just starting to sort out PCI express
ATI radeons right now of course.

Alan

OT: nVidia driver [was: Wish list] -- understanding the GPU market

[Matthew Miller]

On Wed, Jun 08, 2005 at 08:22:33AM -0400, Alan Cox wrote:
>  Its more an inevitability than the reverse. Commoditisation means
> that very
>  soon all the 'must keep secret' IP will be only relevant to real time
>  ray tracing. What happens to the Radeon 9800 when Intel 9xx/VIA/etc graphics
>  on chip are good enough for gamers. Leading-edge video will be for VR nuts. 
Yes -- this is what'll save us eventually. Remember when on-board sound was
a joke and if you actually wanted to listen to something besides blips and
beeps you bought an add-on sound card? Now, pretty much only people actually
working with audio for a professional reason (or hardcore hobby reason) need
to buy a sound card. This day will come with video too.

-- 
Matthew Miller           mattdm(a)mattdm.org        <http://www.mattdm.org/>
Boston University Linux      ------>                <http://linux.bu.edu/>
Current office temperature: 79 degrees Fahrenheit.

OT: nVidia driver [was: Wish list] -- understanding the GPU market

[Nicolas Mailhot]

On Mer 8 juin 2005 15:09, Matthew Miller a écrit :
>  On Wed, Jun 08, 2005 at 08:22:33AM -0400, Alan Cox wrote:
> > Its more an inevitability than the reverse. Commoditisation means that
> > very
> > soon all the 'must keep secret' IP will be only relevant to real time
> > ray tracing. What happens to the Radeon 9800 when Intel 9xx/VIA/etc
> > graphics
> > on chip are good enough for gamers. Leading-edge video will be for VR
> > nuts.
>
>  Yes -- this is what'll save us eventually. Remember when on-board sound
>  was
>  a joke and if you actually wanted to listen to something besides blips and
>  beeps you bought an add-on sound card? 
And this card is widely supported (by Alsa...). Nowadays you can almost
choose a card based on the performance of its chips, not the brokeness of
its closed driver (though for some high-end audio stuff, people still
choose harware based on the quality of its windows driver, on a "sucks
less" basis)

Regards,

-- 
Nicolas Mailhot

OT: nVidia driver [was: Wish list] -- understanding the GPU market

[Dave Jones]

On Wed, Jun 08, 2005 at 08:22:33AM -0400, Alan Cox wrote:

>  Which would be why folks are currently just starting to sort out PCI
> express
>  ATI radeons right now of course. 
FWIW, I picked up a PCI-E Radeon X300 for not-many-beans recently.
It works out of the box with FC4, with no need for proprietory drivers.
The 3d isn't supported yet (See reverse engineering effort I mentioned
in earlier mail), but for 2D its a damn good card.  I never thought I'd
say I'd find a better dual-head card than a Matrox, but this thing
really shines.

		Dave

Stop! (was Re: OT: nVidia driver)

[Miloslav Trmač]

Why exactly does this discussion belong to fedora-devel-list?
Before answering please reread Fedora Project objective 2.
	Mirek