Author: Jeremiah Penery
Date: 19:11:21 05/07/02
Go up one level in this thread
On May 07, 2002 at 15:09:50, Robert Hyatt wrote: >On May 07, 2002 at 13:45:47, Joshua Lee wrote: > >>I figure you also mean the same number of cpu's as a quad would still outgun a >>Dual 2.53Ghz P4. >>Does anyone out there have the last DEC Alpha at 1Ghz? >>That would be a good Battle. > >New Itaniums are also very fast... faster than any 32 bit cpu I have seen >any test results for... I haven't seen SPECInt results for Itanium. I suspect they're not so hot. Even SPECFP isn't to write home about: http://www.aceshardware.com/SPECmine/index.jsp?b=2&s=0&v=4&if=1&r1f=2&r2f=0&m1f=0&m2f=2&o=0&o=1&o=2 You can see the Itanium is behind the Alpha at about the same clockspeed, and is even behind some SPARC and P4 machines. Of course the Power4 has a commanding lead over everything (it also leads in SPECInt). Some things about the Itanium: First of all, it is a wide-issue _in-order_ VLIW processor. The supposed benefit of this is less complexity, but the Itanium is in reality _more_ complicated than the competing RISC architectures. It has full hardware predication, a ton of registers (via a large rotating register file, which may or may not prove advantageous), and lots of execution units. I don't recall how many FP units it has, but I know it has 6 integer units. In theory, it can issue 6 integer operations/cycle, which would make it of course super-fast. The downside is that the compiler has to be _really good_ in order to keep the pipelines even moderately filled. It has to bundle instructions in groups of 6, and if there is no instruction to bundle, it must send a no-op, since it can't take instructions from somewhere else in the codepath like practically all other modern processors (which feature OOO, superscalar architecture, etc.). On a lot of FP code, this can be a win, because it's very predictable, relatively branch-free (i.e., it's very happy to be executed in-order), and can benefit a lot from the extra execution units. Unfortunately for the Itanium, on a lot of integer code that exists, in-order execution is not good for performance. It's simply too difficult to keep all the units filled with instructions on most integer code today when it must be executed in-order.
This page took 0 seconds to execute
Last modified: Thu, 15 Apr 21 08:11:13 -0700
Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.