[Info-vax] RealWorldTech on Poulson

[Johnny Billquist]

On 2011-07-04 23.57, Michael Kraemer wrote:
> Johnny Billquist schrieb:
>
>> So, the HP 755 (which shows the better numbers) gives about 80/150 for
>> int and fp, while the Alphas best is 130/200. That's more than what
>> I'd call "marginally better".
>
> For a CPU which was three years late and
> which already got a lot of hype before it was available,
> a 30% bottom line advantage needing twice the clock rate
> is not very impressive.
> Moreover, I'm pretty sure the best alpha machine
> (a 7000, not a 3000) was way more expensive than the 755.
> So even price/performance-wise they weren't the best buy.

The performance difference with some smaller Alpha is not that big, so 
we still about a 30% advantage. Who cares that the clock rate is twice? 
That is irrelevant. Or are you saying that the clock ticks themself cost 
you something?

>> But the big problems with the SpecINT and SpecFP tests that they are
>> so dependant on compilers that they don't really show CPU speeds that
>> you can compare, but show how good your compiler is at compiling for
>> those tests. And yes, compilers many times had (do they still?) code
>> that specifically detected these tests, and popped out specially tuned
>> code for them. So it's questionable how relevant information from such
>> tests are.
>
> It's way more relevant than clock speeds, because they test
> the whole system: CPU, RAM, OS, compiler, apps, i.e. exactly
> what I buy as a customer. The Spec suite is a well defined
> average of different codes, so it wouldn't be easy to cheat
> with all of them. Moreover, when the vendors submit results,
> they have to specify
> exactly the configuration / compiler options they have used.
> If some vendor cheats, you bet the competition will notice.
> Of course all vendors will tweak their products,
> but since they will do in the same direction,
> the relative ranking will still be significant.

You do know that the SpecINT 92 has long been obsoleted because it 
turned out to not be a good test, partly because it was too easy for 
compiler writers to detect and explicitly handle the test cases? (There 
were other problems too, which caused the tests to be replaced.)

You seem to place way to much value into the Specmark tests results.

>> DEC was the second largest computer company in the world only a few
>> years earlier than the Alpha. To call it a "small" company is kindof
>> misstating facts.
>
> *Relatively* small. It was second to IBM in formal ranking,
> but if one looks at real numbers, IBM was a $50B+ company around 1990
> and DEC was at $10B+x, iirc. So about a factor of 4 or 5 between them.

So the claim that DEC was too small to make something like the Alpha, 
where does that put Sun, HP and Mips?

	Johnny