QCD Performance Summary

• Streaming SIMD (SSE/SSE2) crucial for P4/XEON

• DDR RAM on 800 Mhz as fast as RDRAM

• 15 % loss single mode compared to dual on XEON

• Parallel perf. on 8 x 8 x 16 x 16 local volume up to 500 Mflops per node double prec. (2002 hardware)

• JLAB implementation shows optimum for 44 local lattice for 1, 2, 3-d - > large clusters

• 50 % scaled efficiency of 124 system on 64 ALiCE processors