30 Dec
2017
30 Dec
'17
5:58 a.m.
On Fri, Dec 29, 2017 at 03:35:27PM -0800, Jon Forrest wrote:
Yeah, that's my memory as well. Alpha was fine but it wasn't the bat
it out of the park that DEC thought it would be.
I've worked at Sun and SGI and had some influence on some of their
processor work (Andy B redesigned the memory interconnect on one of
the SparcStation SMP boxes because I showed him data that proved that
we were about to suck; got a processor at SGI cancelled because it
was just beyond hopeless, it was a floating point machine called
Beast and it was being developed as the focus was shifting heavily
from floating point to integer. The stupid thing had pins on all
6 sides - top, bottom, and the edges. The socket, just the socket,
was predicted to cost SGI $1400/each at volume. No CPU, just the
socket. And it had miserable integer performance. Bzzt.)
What I've seen over and over is teams compete against their previous
design rather than the competition. Many, many times I've seen
designers who had barely a passing knowledge of what the other guys
could do. Blew my mind. It was part of the reason I wrote LMbench,
at least that gives you an idea of what the other guys can do.
--lm
Alpha generally
maintained integer/ALU and clockspeed leadership for
most of the '90s
http://www.cs.columbia.edu/~sedwards/classes/2012/3827-spring/advanced-arch…
As nice as the Alphas were, I don't recall any compelling reason we
would have used them if we had to pay for them. In fact, at the same
time Mike Stonebraker and I wrote a grant proposal to Sun to get a
couple of SparcStation 10s to use to port Postgres to Solaris. The
SparcStations were just as nice to use as the Alphas.