• View
• Edit
• History
• Print

20141010Notes

• ran small version of Correlator on 16x16 Hurricane, but results were very inefficient
  • Hurricane cycles were 10x slower than ideal critical path cycles, suggesting that layout is very bad
  • Palmer is going to hand-write Correlator instructions to help identify where inefficiencies lie
  • hand-written code will also give tape-out # tiles, IO sizes, correlator throughput, etc.
• Albert is adding complex multiply instruction to kick-off DSP instructions and hopefully improve performance
• follow up with Brian Richards about ST28nm