How fast GMP can be on Opterons?

Torbjorn Granlund tege@swox.com
22 Apr 2003 23:39:55 +0200


Guillermo Ballester Valor <gbv@oxixares.com> writes:

  Hi,
  
  Because today is the AMD Opteron launch day I'm wondering how fast this new 
  chip can work with GMP.
  
  I see some impressive data at :
   
  http://www.amd.com/us-en/assets/content_type/DownloadableAssets/Optimization_-_Tim_Wilkens.pdf
  
  1) Latency for a 64x64 bits mul: 4 cycles.
  2) A mul can be dispached every cycle
  3) A 64 bit add every cycle
  ....
  
  In an ideal world it could be more than 4 times faster clock by clock than 
  current Athlons, what is remarkable. 
  
I expect it to be more than 4 times faster, thanks to the 8 added
registers, r8...r15.  Perhaps we could reach 2 cycles/limb for
the critical operation mpn_addmul_1.  That would imply an 8-fold
speedup, clock-for-clock compared to Athlon.

-- 
Torbjörn