Tench,
Maybe I wasn't sure what your question was as you had
Foster and Wildfire in the same paragraph. This question
is much more straight forward:
"With all due respect, Rob, I don't think you answered my question. What happens if 4-way Foster proves to be
more powerful than a 4-way Alpha server in 2001?"
I see a window here where that happens, but that is
happening today too (4 processor Xeon/whatever is doing
better versus Alpha on tpcs). However, in a 2001
timeframe EV7 ships (maybe/hopefully) with on-chip L2,
on-chip network switch (CPU to CPU), on-chip memory switch
at a projected 1.5 GHz. Pretty even race if Foster has
all those features too. In other words, Alpha gains more
than parity and decreased cost by leaving out
*complicated* chipsets (we beat that horse a while back)
and maybe even L3 (jury is still out on that, a periodic
fishing question to comp.arch "What about L3 for 21364
aka EV7"? Since it comes up so often, one wonders if
L3 is even necessary with on-chip memory switch and
RDRAM, seems doable to negate L3.)
But think about that for a while and dredge up the IA64
whitepaper:
digital.com
If you check out page 30, you'll note Linley points out
McKinley doesn't even have on-chip network nor on-chip
memory switches...
So that raises interesting issues... among them that
Foster will not only outperform Merced but also
give McKinley a run for its money (with on-chip trickery).
If Foster lacks these features, won't do well against
EV7 and a bonus with EV7 should allow for very good
server prices by leaving out *complicated* chip-sets
and L3 (conjecture, but reasonable).
Since McKinley is feature poor compared to EV7, how do you
suppose it will do performance-wise versus EV7?
Rob |
