[Pc_Support] AMD Fusion going to keep them in the game ...

[Bryan J. Smith]

Without a doubt, Intel has wrestled back the performance computing
crown with its new Core redesign, chuckling the inefficient and
quickly hacked NetBurst (P4).  But AMD has already knocked Intel back
in the Server I/O and is making it a 1-2 punch with Fusion -- keeping
it lower power and higher performance at a much lower price point.

AMD is still the sole x86 compatible vendor with not just a crossbar
switch, but a direct connect system interconnect.  It solved the
coherency issue between CPUs and other components with the Athlon MP
and its EV6 crossbar, and has taken it to the next extreme with
multiple HyperTransport interconnects between packages, memory and I/O
as well as a crossbar inside of multiple core units.

AMD's first experiment with non-CPU devices on its HyperTransport
system interconnect was the HyperTransport eXtension (HTX), used
largely for direct, inter-mainboard connection (using 4 socket base
units) as well as indirect Infiniband clusters.  What really made
everyone take note of HTX was Infiniband, where performance of HTX
Infiniband was 3x better than not only PCI-X 2.0 Inifiband adapters,
but PCIe x8 as well.

>From that, AMD announced Torrenza.  Although HyperTransport is open,
the way its Opteron handles inter-CPU coherency, especially for I/O
processes that share memory like CPUs, was not until Torrenza.  The
idea is that co-processors and other processing capabilities can and
should work directly via HyperTransport links with Opterons as if they
were Opterons, just like Infiniband HTX does for clustering Opterons
well beyond the base 4 or 8 sockets in a mainboard.

So the next move became obvious.  With Intel's influence over the
peripheral market, AMD wasn't going to break through the performance
graphics market other than niche players like Cray, IBM, Quantum3D and
the like, well outside commodity and economies of scale.  With the
purchase of ATI, although risky, AMD now has that.  But AMD is not
merely going to target the high-end market, but is starting with the
integrated one.

Unlike Intel, which relies on the "single point of contention" Memory
Controller Hub (MCH) aka "northbridge" (NB), and puts the graphics
there using a combination of memory hacks and special processor modes,
AMD is putting the ATI GPU right on-die, right on the crossbar
interconnect.  This will bring down price while giving it priority
memory access just like an Opteron -- all at a Sempron price-point,
and a performance-point that makes Intel whimper.

>From the article "AMD-ATI 'Fusion' Full Speed Ahead":
http://www.internetnews.com/ent-news/article.php/3640251

  Analyst Jon Peddie of Jon Peddie Research, said AMD has
  an advantage over Intel with its integrated memory controller.
  "Adding graphics isn't going to be trivial, but it'll be easier for
  AMD than Intel," said Peddie.

  "Intel has to get rid of its frontside bus before it can think about
  integrating graphics, so we don't think they'll even announce
  anything like what AMD is doing until 2008. We think AMD has
  a two-year lead on this kind of configuration."

AMD went through a lot of growing pains with its EV6 adoption,
including all sorts of software incompatibility issues with AGP on
Athlon, especially Athlon MP, when dealing with Intel's "software
hacks" for consistency.  But AMD has moved well beyond them,
especially with Athlon 64/Opteron, and made the transition to PCIe
painless.  These are lessons that Intel either has to learn themselves
in order to compete, and that's not going to be overnight.

Unless Intel is willing to license HyperTransport and, more
importantly, the core CPU designs that make Torrenza possible.  I
doubt it, and Intel is definitely 2 years behind.  They are just now
starting to offer dual front-side busses for Xeon, but it doesn't
offer and is incompatible with PCIe graphics and its software
coherency hacks -- something AMD solved in hardware over5 years ago.
;->