Final Remarks

• Large lattices

  • HMC -Wilson/staggered, domain wall: apeNext, QCDOC and clusters: < 1 $ / Mflops

• Overlap fermions

  • extremely compute intensive on small systems
  • QCDOC scalable to 8 x 8 x 8 x 16 = 8k procs for 163 x 32
  • apeNEXT scalable to 8 x 8 x (4 x 4) = 512 p for 163 x 32
  • Mesh cluster savely scalable to > 4 x 8 = 32 procs for 163 x 32 lattices

• Cluster might further approach application specific systems in the next 2 years as far as scalability is concerned if program runs in cache (stability of cluster will become an important issue)