### Table 4: n25 error rate for 2-way, 4-way, and 8-way classin0c-

1996

"... In PAGE 2: ... For each combination, we built a discrimination rule using LDA on the training speakers, and then evaluated that rule on the test speakers. Table4 shows the total classin0ccation error rate for each taskn2fphoneme set combination. Error rates varied substantially across tasks, but were gener- ally encouraging.... ..."

Cited by 4

### Table 1 reports the approximate number of cycles taken to execute memory fence and atomic exchange instructions. The numbers have been determined with a single-threaded microbenchmark on a 8-way Power4 1.1 GHz system and a 4-way Power5 1.6GHz system. sync, isync, and lwsync are variants of memory synchronization instructions; the com- bination of the lwarx and stwcx read and update a variable atomically. Note that the cycle time of these instructions is not constant and depends on the dynamic context (like the number of pending stores); the cycle times reported in the table demonstrate nevertheless that cost of memory fence instructions is substantial and can be several times to cost of a memory access.

"... In PAGE 2: ... Table1 : Approximate cycle times for memory synchroniza- tion and atomic-read-update instructions on a Power4 and Power5 multiprocessor server. 1.... ..."

### Table 1: Empirical comparison of stability of the two algorithms 4 Conclusion We have demonstrated that the method for computing roots of unity given in a standard pro- grammer apos;s reference guide [3] is unstable. The instability is apparent from the analysis, and experimental evidence shows that the instability is indeed a problem for realistically sized data sets.More importantly, in equation (3) we have given alternative formulas for the computation of the roots of unity. We have shown both by analysis and experiment that the new equations are indeed stable, and provide substantially more accurate results than the previous formulas. In addition, although the complexity of the equations is not a great issue for this problem, the new formulas are computationally simpler than the previous ones. Given these results, it seems that the formulas in equation (3) are the only reasonable way of computing the complex roots of unity.

### Table 3). However, this is still substantially lower than the

1991

Cited by 5

### Table 1 shows that the latency is substantially higher for the

### Table 1: Comparison of p-QR, p-Kmeans, and K-means for two-way clustering

in Abstract

"... In PAGE 7: ...xample 1. In this example, we look at binary clustering. Wechoose 50 random document vectors each from two newsgroups. We tested 100 runs for each pair of newsgroups, and list the means and standard deviations in Table1 . The two clustering algorithms p-QR and p-Kmeans are comparable to each other, and both are better and sometimes substantially better than K-means.... ..."

### Table 1. Combinations produced for TCAS example 2-way 3-way 4-way 5-way 6-way

### Tables 1, 2, 3, and 4 summarize the results. It is seen from the tables that the lowest NMSE and NCD have been obtained using ANF, followed by AMNFG2 and DW-MTMF. Regardless, we have to mention that, in general, the objective criteria described above can serve only in comparative analysis, since they do not take into account the way images are perceived by the human eye; e.g., the measures fail to assess the sharpness of object or region boundaries after ltering. Thus, it is not appropriate to draw conclusions about ltering performance based only on NMSE and NCD, without visual inspection. The ANF approach has provided better visual results than the other methods: noise is substantially reduced, while edges are not blurred. DW-MTMF provided a better visual result than AMNFG2, due to the fact that edges are less blurred by the former; however, the noise attenuation is not as good as in the case of the ANF. Visually, AHMF has outputs comparable to those of AMNFG2, although the latter procedure is better in terms of the objective measures 15

2000

### Table 11 reveals that a cache using the C/NA strategy has the potential to substantially reduce the amount of necessary bus bandwidth (by as much as 64% in the case of su2cor). Looking at this table and Tables 6 and 9, we also see that while the improved static scheme comes fairly close to the limit of bandwidth reduction, the dynamic schemes are not yet approaching the ideal.

1997

"... In PAGE 21: ... However, the table does show that in the optimal case a direct-mapped cache using C/NA has a hit rate that exceeds that of a 2-way set associative cache using LRU and approaches that of a 2-way set associative cache with optimal replacement or a 4-way set associative cache using standard LRU. Table11 shows the average number of bytes needed per data reference for four different cache configurations, while Table 12 contains the reduction in bandwidth requirements in per- centages compared to the optimal MR at the same level of associativity. The same cache param- eters as before were used, as well as the same bandwidth parameters (regular cache miss fetches 32 bytes, C/NA miss fetches 8).... In PAGE 23: ...rn Table11 : Average Number of Bytes/Reference in the Optimal Case ulululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululul Average Bytes per Reference, 8K byte Cache ulululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululul Direct Mapped Two-way SA Four-way SA Fully Associative ulululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululul Benchmark Optimal Optimal Optimal Optimal Optimal Optimal Optimal Program Standard CNA MR CNA MR CNA MR CNA ulululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululul ulululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululululul compress 6.5005 3.... ..."

Cited by 14

### Table 6.2: Words of the English Lexicon mode of operation of the lexicon system, is the parsing of sentences, immediately after the insertion or deletion of some words of the lexicon. In that way, the user will be able to test the results of his changes directly. A menu driven interaction program would be the most appropriate solution. Moreover, productions of sorted or selected (e.g. by a regular expression) listings of parts of the lexicon would be a very helpful feature. Considering e ciency matters, most common words caching is a method that could make substantially faster the \search quot; process of a word. It is observed that 50% of the words that are used in a document, is a small subset (approximately 100-150 words) of all the words of a language ([VAL 88]). Creating a special page with only these most common words and storing this page in main memory, we could decrease the mean search time. We are aware of the fact that two searches are required for every word that does not reside in the most frequent word list (one in the most common word page and one in the lexicon), but the overall performance will de nitely increase, as a 50% decrease of the number of disk accesses could be achieved.

1995

Cited by 1