[Pc_Support] Re: AMD buys ATI, is this good or bad -- AMD has to control the GPU

[Bryan J. Smith]

On Mon, 2006-07-24 at 12:10 -0400, Ray Brunkow wrote:
> http://news.bbc.co.uk/2/hi/business/5210276.stm
> discuss.  is this good or bad?  Will this mean that ATI
> will now start supporting Linux,

ATI has been supporting Linux in its commercial GPU drivers for the last
year or two.  Their chipset business has been more in its infancy, so
support was more limited -- just as nVidia's chipset support under Linux
took a couple of years of availability to pick up.

> or will AMD stop supporting Linux...  neither both, what.

Considering AMD and ATI themselves, like Intel and nVidia, are heavily
Linux internally, they'd cut their own throat.  Understand the 90% of
the EDA world has been 100% Linux since the start of the 21st Century.

Another a key factor in AMD's purchase of ATI is to move the GPU from
the peripheral bus to where it belongs -- the system interconnect.  That
isn't just for gaming, but _high_end_ CAM engineering and precise
(non-lossy math) simulation.  These markets are also heavily Linux.

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.
A "bus" is easy -- because there is *1* point of contention, the
chipset.  Everything must pass through it -- everything must be _shared_
through it.

AMD opened the gate to higher performance when they switched to the
"switched" interconnect of Digital Alpha EV6 -- and learned a lot (the
"hard way") with the Athlon MP.  The Athlon64/Opteron is the _first_
commodity processor with a _true_ switched/mesh interconnect and
on-board I/O MMU -- an evolution from the _on-chip_ (not "on-chipset"
like Intel) AGPgart to maintain coherence between CPUs doing I/O, namely
AGP video.

Today, the I/O MMU maintains coherency between all processors on the
HyperTransport partial mesh.  HyperTransport processors also have their
own _I/O_ affinity -- I/O is physically (and logically) _localized_ to
each processor.  That's why InfiniBand over HyperTransport is 3x faster
than PCIe x16 -- and scales _linearly_ for multiple HyperTransport
interconnects from multiple CPUs.

Now imagine that for video cards!  Especially high-end CAM solutions
that are almost always running Linux -- as well as entertainment.

Intel can't do it.  That's why we have PCIe, yet another peripheral bus.
The GPU is a CPU.  It needs to be coherent with CPUs, on the same
interconnect.  AMD doesn't have the marketing muscle to get GPU vendors
to do HyperTransport.  So AMD finally just bought a GPU vendor.

I don't know if they'll have the money to do so.  That's the scary part.
But they're doing it regardless.


-- 
Bryan J. Smith          Professional, technical annoyance
mailto:b.j.smith at ieee.org    http://thebs413.blogspot.com
---------------------------------------------------------
The world is in need of solutions.  Unfortunately, people
seem to be more interested in blindly aligning themselves
with one of only two viewponts -- an "us v. them" debate
that has nothing to do with finding an actual solution.