[Pc_Support] Re: AMD buys ATI, is this good or bad -- this will *ELIMINATE* ATI's proprietary kernel driver

[Justin M. Keyes]

On 7/25/06, Bryan J. Smith <b.j.smith at ieee.org> wrote:
> On Tue, 2006-07-25 at 09:30 -0400, Bryan J. Smith wrote:
> > Intel's current platform "bus" interconnect makes this _impossible_.
> > That's why Intel has not moved past using AGP and, now, PCIe -- both I/O
> > interconnects -- as video interconnects.  It uses 100% _software_ to
> > maintain CPU to GPU coherence, because it's CPUs _lack_ anything else.
>
> As I regularly point out, this "100% _software_ to maintain CPU to GPU
> coherence" is virtually 99% of the reason why ATI and nVidia have
> proprietary kernel drivers**.  Unlike other I/O peripheral cards that
> merely do "DMA" transfers to/from the system memory, video cards do,
> what Intel calls, "Direct in-Memory Execution" (D[i]ME).  In other
> words, it acts like a CPU -- using memory _directly_.
>
> [ **NOTE:  The proprietary kernel drivers should _not_ be confused with
> the proprietary XFree/Xorg (X11) driver and related LibGL library and
> resulting GLX (OpenGL over X11) support.  Those drivers are _allowed_ by
> both the XFree/Xorg license.  The kernel driver is used for memory
> _performance_ at the kernel level, which is more questionable. ]
>
> The code that both ATI and nVidia use is an Intel _trade_secret_.
>
> For the longest time, Intel even considered the AGPgart to be a trade
> secret.  Back then, only nVidia (not ATI) has chipsets, which is why the
> nForce AGPgart wasn't part of the stock kernel.  Once PCIe came out,
> Intel stopped considering its AGPgart as a trade secret, so nVidia was
> able to add the nForce to the stock kernel.
>
> Now I'm sure you're saying, "Wait as second Bryan!  Intel's own chipset
> AGPgarts are in the kernel, and Intel doesn't have proprietary kernel
> drivers!  So you're wrong!"
>
> And my response is, as always, "Have you seen the performance of the
> Intel drivers on Linux versus Windows?"  Yeah, that's right, it _sucks_.
> Intel's AGPgart and [almost total lack] of a GPU support driver _kills_
> performance on Linux.  Memory coherency is _everything_.
>
> So what's this have to do with ATI-AMD?
>
> AMD's I/O MMU and inter-CPU coherence is in the CPU.  It's _hardware_,
> not a flaky set of software and hacks, like Intel.  So not only do we
> get a _stable_ way for the GPU to access system memory, because it's now
> on the system interconnect instead of a peripheral card trying to act
> like a local CPU with software hacks -- but it will *ELIMINATE* ATI's
> proprietary kernel driver.
>
> I've said it before and I'll say it again -- Intel has been the problem!
> Not ATI.  Not nVidia.  *INTEL*.  And this move by AMD will *ELIMINATE*
> it, at least for ATI.  ;->
>
> nVidia will have to decide if it wishes to support AMD's new HTX GPU
> standard.  And although it owns ATI now, I believe it will work with
> nVidia to do so.  Because freezing Intel out of the high-end GPU market,
> or forcing Intel to _redesign_ their _entire_ platform to support it
> (which will take at least 18-24 months -- possibly longer for stable I/O
> MMU support which AMD took a bit to learn with the Athlon MP and AGP) is
> more of a marketing strategy than "beating nVidia" with its ATI arm.

Both emails on this subject have been very enlightening...

-- 
Justin M. Keyes