Elmer,
<Intel is pinning its hopes on McKinley, its second IA-64 processor, to establish the new architecture as the sole
performance leader. We expect that chip, due in late 2001 at clock speeds exceeding 1 GHz, will open a
performance gap of 2030% over the fastest RISC processors. In 2002, a 0.13-micron derivative of McKinley,
code-named Madison, will take performance to greater levels. >
Yeah... a real chucklehead, eh? Nice to have a sense of humor. McKinley pilots
late 2001, but ships 2Q2002? Maybe? But that is a nit at this point. The statement:
"will open up a performance gap of 20-30% over the fastest RISC processors." Yes, that
is a laugher. Maybe he is forgetting all about Power4? No. Like many, he is ignoring it. Power4
will deliver tpmC that will be very scary indeed, look at its cache and memory and I/O numbers:
realworldtech.com
chips.ibm.com
Read that PDF and read what we know about McKinley. McKinley is in a
different segment.
In that article , Keith D. states:
"The great unsolved mystery is why Intel/HP and IBM arrived at such
polar-opposite solutions, Intel and HP [with IA64] have obviously
focused their efforts on exploiting single-thread ILP, with less concern
for TLP or memory bandwidth. At the opposite extreme, IBM has focused on massive
memory bandwidth and TLP but paid only moderate attention
to ILP. Intel obviously believe there is enough latent ILP lying
around [yeah, and only HPIntel knows where it is hiding!] to justify
a departure from the most dominant architectural franchise in the
history of mankind. Intel says it has made the switch to a new
ISA [IA64] at this time to give it a solid platform to which it can
later add TLP and high-bandwidth interfaces. It believes that others will
eventually be forced to make this same ISA transition to avoid leaving
a wealth of parallelism on the table.
IBM, on the other hand, clings [ you left out: " in desperation " ] to a far
less pervasive ISA, seeing little rationale for more minor tweeks. IBM says that
memory bandwidth is the limiting factor today and predicts that it will
only get worse over time. The company believe that the parallelism achievable
with superscalar, multithreading, and multiprocessing can saturate any practical
memory system, now and until quantum dots replace transistors. Thus, the whole
issue of the ISA is simply a moot point."
Something is obviously amiss; both camps cannot be right.
---
The rest is good reading too. Compaq with EV8 + SMT is pursuing the
same path as IBM and took a long hard look at VLIW, prior to HP/Intel
plunging in VLIW and Digital/Compaq rejected it, see the bottom of
page 3:
alphapowered.com
EV7 with 44.6 GByte/sec memory and I/O bandwidth before EV8....
Interesting ride coming up.
Rob |
