Jesper Frimann (jesperfrimann@gmail.com) on 8/27/09 wrote:
---------------------------
>someone (someone@somewhere.com) on 8/25/09 wrote:
>---------------------------
>>mpx (mpx@nomail.pl) on 8/25/09 wrote:
>>---------------------------
>>[..]
>>>Now, with enough time to
>>>spend on Power7 architecture they are back with a superior design.
>>
>>Technical tidbits disclosed so far are intriguing
>>but that statement clearly remains to be seen.
>>
>>In the prelease hype before Power6 some folks who
>>should have known better got caught up in high GHz
>>fever and forgot that there would have to be design
>>tradeoffs that IBM would go far out of its way not
>>to mention. Heck, IBM still won't even hang a TDP
>>number on it. Power6 turned out to be remarkably
>>inefficient in its ability to turn frequency, issue width,
>>and bandwidth into application level performance.
>
>Well, being an architect I don't get to do much hands on anymore. But before my
>vacation I was involved in a performance tuning project on a POWER6 machine. A real
>bitch project with 300+ hours of work in less than a month.
>It turned out that I was the only available guy with serious big system performance
>tuning experience, so the old rusty man had to give it a go.
>
>And I must say I was quite stunned by how fast, the now old, power 570 (power6 4.7GHz) we were running on was.
>Simply a monster that ate everything that we threw at it.
>
>I got the same feeling of speed and horsepower that I got when I tried my first
>POWER4 p690 when it was brand new, the first ES45 or the very first PARISC hp9000 730 'Snake', workstation I tried.
>
>So you might not like the engine, but damn the truck rocks.
>
>// Jesper

I am not saying that Power6 isn't a fast processor. The
real question is given all the basic inputs - clock speed,
issue width, and memory bandwidth - why it isn't even
faster. Of course $, Watts, mm2 of Si are the real world
costs so if IBM can pile up inputs to burn within those
constraints then how can you complain about Power6
being very inefficient with them?

My real cautionary objection to the heavy breathing over
Power7 Hotchips disclosure is to pay as much attention
to what IBM isn't saying as to what it is. IBM is extremely
oleaginous when it comes to hyping its engineering and
products to the technology press. I do remember some
folks here at RWT arguing based on the initial megahyped
Power6 disclosures that it would perform like a Power5
scaled up in frequency and vociferously denying there
would have to be yet unspoken design compromises to
reach those frequencies. Pity the sceptic who suggested
IBM might have reduced, or god forbid, eliminated OOOE
capabilities to achieve the necessary logic simplication
per pipe line stage in Power6 to clock that fast in 65 nm. :-)