🚀 go-pugleaf

In article <1666@awdprime.UUCP> @cs.utexas.edu:ibmchs!auschs!woan.austin.ibm.com!ron writes:
>
>In article <1653@awdprime.UUCP>, ron@woan.austin.ibm.com (Ronald S.
>Woan/2100000) writes:
+|> I must have missed something here... What integer unit parallelism?
+|> From what I understand, the integer unit is a single pipeline (5?
+|> stages), so it can never do better than one instruction/cycle. To get
+|> five/cycle, you need two branches, float add, float mult, and integer
+|> operation. Only an amazing compiler can schedule an application to use
+|> this frequently.
+
+Some people within IBM have asked me to clarify my statements (I
[stuff deleted]
+branches with respect to integer operations; however, with a single
+integer operation pipeline, you will never see more than one integer
+operation/cycle. Also, it would take an amazing compiler to schedule
+five concurrent operations regularly for any imaginable typical
+application to maintain the peak execution figure that some people
+have been quoting.

My understanding is that there are 3 ALU's, each with it's own
pipeline, and that is where the 5 ops/cycle peak is (i.e. 3 int
ops, 1 fp mult and 1 fp add).  Of course, you don't often get
more than 1 int opn at a time, as is revealed by the discrepancy
between the fp and int benchmarks.

+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+=+
Greg Frazier				"Big A, little a / What begins with A?
frazier@CS.UCLA.EDU			Aunt Annie's Alligator / A ... a ... A"
!{ucbvax,rutgers}!ucla-cs!frazier			_Dr._Seuss's_ABCs_