[geeks] Carly's Gone!!!

Francisco Javier Mesa-Martinez lefa at ucsc.edu
Thu Feb 10 12:08:32 CST 2005


On Thu, 10 Feb 2005, Jochen Kunz wrote:

> > > high memory bandwith. That is the reason why intel put huge (and
> > > thus expensive) caches on the chip.
> > Well that is the same problem with almost any new CPU isn't it?
> Sure. I have some old Alphas. The low end models have 64 bit memory bus,
> the high end machines 256 bit wide memory... Alpha eats _lots_ of memory
> (bandwith). But from what I heared Itanium is even worse then Alpha in
> terms of memory bandwith.

Well memory bandwidth requirements are rather hard to compare accross
architectures, so I have no idea how they came up with that claim. The
rule of thumb pretty much is: there is never enough memory bandwidth :) so
whether it is Itanium or Alpha or Power they all need as much BW as they
can get. However, Itanium is still an in order design, so it really can
not tolerate memory (spec data) latencies as well as the very aggressive
21264 (and 364) out-of-order cores. Same goes for the POWER4/5 when
compared to UltraSparc, the US needs globs of cache to perform decently
(and even then it can hardly keep up with other modern competitors) since
it is not out-of-order whereas the newer POWER micros are.

> > To be fair neither SMT nor differential high speed serial channels
> > were invented because of alpha.
> Interresting. Alpha was the first CPU that brought SMT and high speed
> serial CPU interconnects to my attention. There where other (cc)NUMA
> designes before, but to my knowledge Alpha was the first CPU that
> implemented it on the CPU chip.

Well, MIPS was ahead of Alpha at least when they were bringing up Beast et
al (the core that was being developed after the R10K, what would have been
the R20K) they had memory controllers on chip to move the NUMA into die
from their custom routers and stuff. Basically they were moving some of
the NUMA-link logic into the chip itself, SGI could have now super dense
Origin like machines and that was in the late 90s! IBM had played with
multithreading before Alpha, however the people who were developing SMT at
UW got a nice grant from Compaq so they used Alpha as the base of their
studies in the simulator. I believe the SMT Alpha silicon was behind intel
first hyperthreading chips....

Still the 21264 and 21364 are amazing microarchitectures....



More information about the geeks mailing list