That's not what they were talking about. Adding SMT to a core and then adding more execution resource to the same core allows either thread access to all the execution resource, but increases overall utilization.
Unfortunately the article is full of crap because IBM's POWER5 design did not do what they were saying -- it does not add that many more execution resources. IBM's presentation at MDF came 3 months after The Register published that "article".
That's not what they were talking about. Adding SMT to a core and then adding more execution resource to the same core allows either thread access to all the execution resource, but increases overall utilization.
I agree with you, but my point was that the only way they're doubling the execution resources is going dual core, instead of doubling the execution units on the one core itself. I believe IBM estimated that for example the integer units were only at some 20% utilization on "average", so a wider core is actually not needed for increased performance. So the point of SMT is to better make use of the units already there. Or am I totally off base here?
Comments
one more AltiVec, two more floating point units, and two more integer units.
*Droooooooooooooooooollllll*
/me inserts missing Homer-look-a-like drooling smiley
Jimzip
Originally posted by T'hain Esh Kelch
*Droooooooooooooooooollllll*
/me inserts missing Homer-look-a-like drooling smiley
It's called 'dual core'. :-P
Originally posted by Zapchud
It's called 'dual core'. :-P
That's not what they were talking about. Adding SMT to a core and then adding more execution resource to the same core allows either thread access to all the execution resource, but increases overall utilization.
Unfortunately the article is full of crap because IBM's POWER5 design did not do what they were saying -- it does not add that many more execution resources. IBM's presentation at MDF came 3 months after The Register published that "article".
Originally posted by Programmer
That's not what they were talking about. Adding SMT to a core and then adding more execution resource to the same core allows either thread access to all the execution resource, but increases overall utilization.
I agree with you, but my point was that the only way they're doubling the execution resources is going dual core, instead of doubling the execution units on the one core itself. I believe IBM estimated that for example the integer units were only at some 20% utilization on "average", so a wider core is actually not needed for increased performance. So the point of SMT is to better make use of the units already there. Or am I totally off base here?