> [ re: pc164 vs celeron benchmark ]

Hmm, random comments:

	* I think the out-of-order execution (NOT invented by Intel) actually
	  IS a big difference. That plus the register renaming (needed more
	  on x86 than alpha, but it works around bad compilers effectively)
	  was Intel's main speed strategy. They investigated the return on
	  silicon area and went with mainly those two optimizations for
	  rather carefully considered reasons.

	* As noted, x86 gcc vs alpha gcc is an apples-vs-oranges test

	* In fact, Celeron vs PC164 is not completely fair, as the current
	  alpha offerings are ev6-based systems. These, BTW, do both
	  out-of-order execution and register renaming. My 264dp will
	  beat my celeron even on the (apples-vs-oranges) gcc test:

		2.8  264dp	time make tcp_input.o
		4.4  Celeron

	* Things are different on floating point..

		SpecFP95

		10 	typical x86 PeeCee on SpecFP95
		16	top-of-the-line x86 PeeCee
		52	500 MHz EV6
		67	667 MHz EV6

	* On Whetstone... (Wow, people still run Whetstone?!) With similar
	  library support the alpha would probably soundly clobber the
	  Celeron here, but NetBSD doesn't have similar libraries.  Whetstone
	  is (unlike real programs) heavy on transcendentals, and NetBSD
	  uses J.T. Conklin's hand-written assembly for these on i386 but
	  the MI SunPro library written in C on alpha.

 	  Looking at a vastly more respected FP benchmark (Whetstone was
	  always something of a joke) the SpecFP95 values for PeeCees
	  and EV6 alphas are widely disparate as noted above.


The Intel processors are in fact quite fast, so it's hard to beat them.
All of the P6-series of Intel chips are 3-way superscalar for integer ops,
plus 1-way for floats, whereas the EV5 alphas were 2+2. This is part of the
difference. Intel eventually caught up with the EV5 generation, but now they
have been leapfrogged again. If you wanted to compare performance-per-dollar,
Intel wins, but it's still apples-and-oranges: e.g., alpha is 64-bit.

	ross.harvey@computer.org