====================================================================== IBIS INTERCONNECT MODELING AD HOC TASK GROUP MEETING MINUTES AND AGENDA http://www.eda.org/ibis/adhoc/interconnect/ Mailing list: ibis-interconn@freelists.org ====================================================================== Next Meeting Wednesday, July 15, 2009 8 AM US Pacific Time Telephone Bridge Passcode 916-356-2663 5 257-2357 (for international and alternate US numbers, contact Michael Mirmak) LiveMeeting: https://webjoin.intel.com/?passcode=2572357 Agenda: - Call for patents - Opens - Continue review of sparse matrix proposal language ====================================================================== Minutes from July 1: Attendees: ---------- (* denotes present) Agilent - Radek Biernacki*, John Moore, Ken Wong Ansoft - Denis Soldo Cadence Design Systems - Terry Jernberg, Brad Griffin Cisco Systems - Mike LaBonte Green Streak Programs - Lynne Green Hewlett-Packard - Rob Elliott Intel - Michael Mirmak* Mentor Graphics Corp. - John Angulo, Vladimir Dmitriev-Zdorov* Micron Technology - Randy Wolff Sigrity - Sam Chitwood, Raymond Y. Chen, Tao Su, Brad Brim* SiSoft - Walter Katz* Teraspeed Consulting Group - Bob Ross* ======================================================================== No patents were declared. Michael presented his proposal for several keywords to express sparse matrices. He also reviewed comments and examples from Walter, showing alternatives. Walter noted that zeroes don't need to occupy physical space in the file. The zeroes can simply be excluded or not mentioned. Radek suggested this might be confusing. An alternate structure would omit listing any elements that were zero; Brad agreed, suggesting that a list is all that is required. Radek added that a list of the non-zero matrices in a single vector would eliminate the need for extra keywords. Michael created a quick example showing [Unique Values] as an example keyword, in the following structure: [Unique Values] (5, 230) : 25 -120 (900, 25) : 75 80 (7, 999) : 1 -23 where the first pair, in parentheses, corresponds to a row and column location while the data after the colon correspond to magnitude and angle data points. Vladimir asked whether each frequency would require a label. Radek noted that the data given would correspond to the largest unique matrix in the file. Brad responded that this might generate too big a matrix. Michael added a question regarding whether frequency-distinct values would be required. Radek suggested this would be a very special, hypothetical case. Bob and Vladimir noted that duplication and zeroes would occur in the matrix because that duplication or zero-valued data was structurally required, not to describe incidental occurrences. Radek reiterated that the mapping need only list non-zero elements, where the header specifies which row, column combinations are non-zero. Brad explained an example where, in a 4x4 matrix, duplicated, upper-half and zeroed data could result in a 16 element matrix only containing 3 data pairs. A format that only includes non-zero elements would not provide this level of compression. Radek suggested this could be addressed and took the AR to provide an alternative format with examples. Walter suggested that most of the alternatives suggested for sparse matrices were equivalent. He added that [Transfer Function] might be a suitable alternative keyword to use. Michael asked whether the relationships in a sparse matrix would truly apply for all frequencies and whether the resulting matrices would always be the same size. These issues might make reuse of [Network Data] difficult. In addition, this proposal doesn't address swathing. Vladimir asked whether this makes any difference. Walter and Bob both suggested that swathing would be possible using this format. Bob added that ignoring zero-valued elements would be preferred, but that any proposal would still need to address upper/lower matrix mapping and mixed-mode treatments explicitly. ARs Radek: show vector-based alternative to multi-keyword proposals.