Tench,
Rob, <With 256 CPUs what is the worst case hop? When
you plot all combinations of hops worst case is a very
low percent. Average latency is very good.>
"Sure, if all you have is one active CPU and an idle interconnect fabric. Toss in 255 other CPUs doing remote accesses simultaneously and you'll run into real scalability issues."
Come on.. you're throwing in a mathematical possibility that
is so remote you have better odds of winning a $100 million
lottery pot. Their are 4 router paths per CPU so many
remote memory ops can be active at the same time. No doubt
some very difficult math here but taken into account.
"If you can show me one 256-CPU system that can be done in a UMA/RUMA/SKUMA way, I'd sure like to see it."
Are you saying it won't happen, PERIOD?
<Nice too is aggregate memory bandwidth, 12.8 GByte/s direct
attached, 25.6 GByte/s routed.>
"You sure about that? I've read each EV7 processor has 6.4 GB/sec memory bandwidth. And adding the "routed" bandwidth is mere marketing fluff. If you want to hype up the EV7, you might want to stick to facts, like the fact that the bandwidth is per-processor, unlike any Intel-based server."
Sure. Read slide 22 of 24 at alphapowered. It states:
32 GB/s interconnect bandwidth. One other thing you
overlooked, each EV7 has two memory controllers and the
slides also state 12.8 GB/s memory bandwidth.
Why 32 GB/s? They are adding in the 6.4 GB/s I/O bandwidth.
I am just focusing on memory bandwidth. After all, seems
McKinley will be competing with EV7. And I heard McKinley
has 3x the bandwidth of Itanium I.
Rob |
