[Pc_Support] AMD Readying PCIe HyperTransport Tunnel ...

Bryan J. Smith thebs413 at gmail.com
Wed Oct 4 11:12:38 EDT 2006


AMD is readying a new AMD 8000 series PCIe HyperTransport Tunnel for
the workstation and enthusiast market.  I've been trying to find out
more information on the IC, but there's little other than block
diagrams of the solutions -- many of them _inaccurate_.

Most system ICs used in AMD platforms these days are a single or
dual-chip combination -- one with a set of 17-24 PCIe channels and
peripheral logic, another with additional peripheral logic, possibly
more PCIe channels.  Other than ViA, who -- like Intel -- is still
leveraging 533MBps-2.0GBps peripheral PCI logic as a "poor man's
system interconnect" -- ATI, nVidia and SiS are all using 1.0-8.0GBps
HyperTransport for true system interconnect.  But despite the
flexibility of HyperTransport, most of these designs are still 1 and
optional 2-chip only configurations.

This new AMD 8000 series HyperTransport Tunnel will allow PCIe
channels to be placed in-line anywhere, to any processor, for
workstations and enthusiast game systems -- much like the AMD8131/8132
is used for PCI-X 1.0/2.0 is used in servers.  So it won't be a
surprise to find not only dual, but even quad PCIe video cards in
higher end systems.

Despite the gross ignorance of most of the enthusiast community, this
HyperTransport Tunnel IC will work with _any_ AMD platform.  The
question is, how well?  If you ask the enthusiasts, their articles
bitch'n moan about AMD not offering it on the Socket-AM2/940.  Little
do they stop to realize why it won't be sold there.

On a Socket-AM2/940 platform, with limited HyperTransport
interconnects, the solution won't offer much over an nVidia nForce
SLI32 or 590 platform -- as all I/O still throughs through a single
HyperTransport link to one processor.  You're going to get the same
performance with those -- because the limitation is at the processor,
_not_ ICs.

The nForce 4 Professional, designed for Socket-940 (not AMD2) and
newer Socket-F/LGA-1207 using the 2200 on one processor with the 2050
on another, is a better solution -- because it relies on multiple
HyperTransports, one from each CPU.

So the power of this new addition to the AMD 8000 series is when it is
used in the same way.  Each AMD 8000 PCIe HyperTransport tunnel has
its own connection to its own CPU, possibly two ICs per CPU (in a
4-way GPU configuration to 2 sockets for processors), using
Socket-F/LGA-1207 Opteron 200 or 800 series.  This can't be done with
the Socket-AM2/940 solution.

I'll draw some diagrams and post to my blog later in the week that
explains its better graphically, especially for the ignoramous
enthusiasts sites that "just don't get it."



More information about the Pc_support mailing list