[log in to unmask] wrote:
 
: Last week's Computerworld had a front page article talking about the P7
: chip from Intel.  It is indeed the first chip based on the new HP and Intel
: architecture code named "Tahoe".  Intel will apparently release a 32bit
: version in late 1997, to be followed by the real 64bit version. the 32bit
: version will be an Intel-only chip (HP is apparently not going to bother
: with a 32bit version, but Intel is worried about the risk in jumping
: directly to 64bits).  Both versions will run on 2.5V.  The 32bit version
: will allegedly get 1,000 SPECint92, and the 64bit version 1,500 SPECint92
: performance.  The "Pentium Pro" at 150Mhz is supposed to around 250
: SPECint92 or so, so we are talking about a 4-6x increase in speed in only
: a two year period.
 
: The new architecture will not be able to execute either PS-RISC or x86
: instructions directly, but will supposedly be able to emulate both
: significantly faster than today's processors can run them natively.
 
: This kind of performance is dependent on a new generation of compilers
: that can take advantage of Tahoe's VLIW architecture.  Much of the job
: of running your program moves from the CPU back to the compiler.  This
: makes the quality of the compilers / code generators essential to the
: performance of the system, much more so in fact than for a traditional
: RISC architecture like PA-RISC.  Since writing quality software (like
: compilers) has proven to be so much harder than developing fast CPU
: chips, it will be interesting to see if HP can really pull this off.
: It also may mean that either all third-party compiler writers are out
: of business, or everyone will be using an HP developed code generator,
: which means that the performance you get and the bugs you see may be
: the same regardless of whether you use an HP, Microsoft, or Borland
: compiler.
 
  HP is suppling both the front end and the back end of the compilier.
  I find it very interesting what different front ends do to the back
  end.  The front-end can tie the back-end into knots.  The C++ front
  end is a port.  ( I can't find out right now if it is OK to say who
  from, but you probably all ready know anyway. )  So even if Microsoft
  and Borland use the same back end, there could be different bugs.
  Also since back-ends are as good as the features it supports, there
  could be performance differences.  For example, in early Spectrum days,
  MPE was getting lousy performance.  We looked at the code and found
  out the long pointer usage was unoptimized.  We went to the optimizer
  people and ask them to optimize long pointer usage.  They said: Oh,
  we didn't know, UNIX doesn't do that, but we'll do it.  It just was
  a feature they didn't know that people used a lot.  This could probably
  happen with the next back-end.
 
: It's gonna be interesting :-)
 
You're right, but it keeps me working.
 
Rick