Viability of Itanium-- was Intel found to be abusing market power in Japan

On Thu, 24 Mar 2005 07:04:09 -0500, Yousuf Khan wrote:

keith wrote:
Opteron is 8-way capable today, but the mechanics get a little
problematic. There is no reason to assume that a dual-Opteron can't reach
8-way on a 4-W board. If you're thinking of a 16P system, AIUI, it's not
going to happen. There would have to be changes to the coherent HT links
for that to happen. I don't see it, but...


Or an external chipset, like Horus, or Serverworks/Sun's new chipset.

Okay, but we aren't talking about the same thing anymore. AIUI, the AMD
NUMA architecture is sufficiently UMA to be UMA as long as the local
interconnects (hypertransport) are used. Once another bridge is used,
this breaks down and we really have a NUMA system. No?

--
Keith
Yousuf Khan

keith wrote:
Or an external chipset, like Horus, or Serverworks/Sun's new chipset.


Okay, but we aren't talking about the same thing anymore. AIUI, the AMD
NUMA architecture is sufficiently UMA to be UMA as long as the local
interconnects (hypertransport) are used. Once another bridge is used,
this breaks down and we really have a NUMA system. No?

Well, even if they used bog-standard Hypertransport to try and go beyond
8P, they may find that the number of hops makes a difference to the
latency that would take it outside of the domain of Sufficiently UMA.

However, there is some talk that the next generation Opterons will be
able to connect to 32-processors all by itself. That would likely
indicate that AMD is planning to increase the number of external HTT
channels per processor.

Yousuf Khan