Under integer tests SPEC CPU2000, Presler loses a little more high-frequency Prescott only in problems 186.crafty (up to-2. %), 252.eon (up to-4. %) and 300.twolf (up to-2. %) a little. In the majority of the others, irrespective of a variant of optimization of a code insignificant advantage, as a rule, is observed. Exception is made with a problem 181.mcf in which it is reached from 16. up to 20. % of advantage above kernel Prescott. Apparently, (hence, increased PSP) that increased frequency FSB here affects platform Presler. At the same time, this separate problem brings the small contribution to the general estimation of productivity SPECint_base20 as on last 65-nanometer the kernel wins against the 90-nanometer of analogue only slightly (up to 2. %).
Among tests with material numbers, only the problem 168.wupwise least unequivocally behaves, showing significant disorder in relative sizes of speed depending on a variant of rugged notebook review optimization of a code, and both in positive, and in negative the parties from 10. up to-20. %. Other problems behave much more predictedly and or are practically tolerant to a choice of the processor (177.mesa and 200.sixtrack), or receive osch'utimoe advantage on Presler (the greatest at 171.swim and 179.art, up to %). Advantageous there is also general rating SPECfp_base20 on the processor kernel Presler, showing up to 8. % of advantage above Prescott, even despite of a little bit smaller clock frequency. On this parameter size PSP can affect a platform with processor Pentium Extreme Edition 965, and micro-architectural features of kernel Presler, like presence of hardware presample of data at a level of a L2-cache of the processor as about l'shaya. Also it is possible to assume, that kernel Presler possesses the best productivity at operations with material rugged notebook review numbers (and as in case of x87-, and a SSE/SSE2/SSE3-code) in comparison with ravnochastotny'm, and even a little more high-frequency kernel Prescott. As processor Pentium Extreme Edition 9 is dvuh`yaderny'm, we shall estimate now efficiency of parallel start of several spears of problems SPEC CPU2000, using for this purpose the metrics rate. Thus as the standard we shall use the sizes received in this metrics with quantity of spears, equal 1, and as the basic measurement with the quantity of spears equal (to quantity of the logic kernels which are present at the given processor).
All integer tests receive a gain in speed at use of the second kernel and technology Hyper-Threading. Of the general picture drops out only 181.mcf in which 11- % of a gain are observed only. Apparently, it is connected with insistence of this problem rugged notebook review already mentioned above to PSP as transition to two kernels is not accompanied by so multiple increase PSP (in view of the general for both kernels of the system trunk). In the majority of other problems the gain in the speed, testifying about efficiency not only dvuh`yadernosti, but also technologies Hyper-Threading (since theoretically maximal gain from use of the second kernel in this case makes 1 %) is observed more than 1 %. Most differ in this plan of a problem 164.gzip and 256.bzip2. By integrated estimation SPECint_rate2000, the gain in speed from the second kernel also exceeds 1 % and makes 104-1 %.
If the rugged notebook review integer tests considered above all without exception differed a gain in speed at transition to a 4-line mode of testing among problems with material numbers is not only those, not showing a gain (171.swim in the majority of variants), but also losing in speed at start of spears (179.art, showing up to % of decline of productivity). The majority of other problems shows a significant gain in speed which in some cases also appears over 1 % (200.sixtrack and 301.apsi). By integrated estimation SPECfp_rate2000, the 4-line mode rate start of the test on dvuh`yadernom processor Pentium Extreme Edition 9 also appears advantageous on 54- % in comparison with similar one-line start. Processor Intel Core Duo T25 Similarly, we shall begin consideration of results with absolute parameters of rugged notebook review productivity of tests SPEC CPU2000. Table 2.1. SPECint CPU2000, Core Duo T2500, 2. GHz Without wholesale.-QxK-QxW-QxN-QxB-QxP 164.gzip 10 11 11 11 11 11 175.vpr 12 13 13 13 13 176.gcc 18 18 18 18 18 181.mcf 13 13 13 18 18 18 186.crafty 12 12 14 14 14 14 197.parser 9 9 9 9 9 9 252.eon 17 17 18 19 20 19 253.perlbmk 16 16 16 16 17 16 254.gap 16 15 16 16 16 16 255.vortex 256.bzip 12 11 11 11 12 12 300.twolf 16 20 20 20 20 20 SPECint_base20 13 14 14 15 15 15 Table 2.2. SPECfp CPU2000, Core Duo T2500, 2. GHz Without wholesale.-QxK-QxW-QxN-QxB-QxP 168.wupwise 12 12 12 12 13 17 171.swim 17 19 19 19 19 19 172.mgrid 9 12 11 11 12 11 173.applu 10 rugged notebook review 10 10 12 11 12 177.mesa 12 15 16 15 15 16 178.galgel 17 26 26 30 30 30 179.art 26 26 28 29 29 28 183.equake 16 15 15 15 15 17 187.facerec 14 16 16 16 16 16 188.ammp 9 8 8 8 8 8 189.lucas 15 14 16 16 16 16 191.fma3d 10 10 11 11 11 11 200.sixtrack 4 4 5 5 5 5 301.apsi 10 10 10 10 10 10 SPECfp_base20 12 13 13 14 14 14 Owing to realization of support of instructions SSE in processor kernel Yonah it appears possible to use a variant of optimization of a code-QxP, traditionally concerning to the processor kernel Prescott initially possessing support of the above-named instructions. For not rugged notebook review clear reasons, calculation of final estimation SPECint_base20 on processor Core Duo T25 in this case does not pass (in spite of the fact that in this series of tests and in the previous series on Pentium Extreme Edition 9 was used the same binary code), therefore the corresponding estimation has been counted up manually, as an average geometrical by results of all podtestov. Apparently, the variant-QxP in integer problems on the average practically does not concede to a variant-QxB, to the most suitable for all previous processors of class Pentium M. And in tests with material numbers the variant-QxP and at all appears advantageous in comparison with-QxB and all other variants of optimization of a code. Efficiency of the last as in integer, and material tests decreases in a number-QxB-QxN-QxW-QxK without optimization. For consideration of the received results in relative sizes as the sample of rugged notebook review comparison we shall choose the investigated processor Intel Pentium M 7 also possessing rather close characteristics with the investigated processor (frequency of a kernel of 2. GHz against 2. GHz, volume of a L2-cache of MB, being in full volume L at Pentium M and effective volume L at use only one of kernels of processor Core Duo).
In comparison with Pentium M, the majority of integer problems SPEC CPU20 or is not received with advantages, or lose in speed on less high-frequency processor Core Duo a little. Exception make already known on processor Pentium Extreme Edition considered earlier 9 problem 181.mcf, receiving up to 30. % of rugged notebook review advantage, obviously, due to higher throughput of the processor trunk and memory on a new platform, and also 254.gap (up to 33. % of advantage), not showing a clear advantage on Presler and, hence, in this case receiving a gain not only because of higher PSP. Nevertheless, general estimation SPECint_base20 in relative sizes again is only slightly more advantageous (up to 3. %) on Core Duo 2. GHz in comparison with Pentium M 2. GHz.
Is much better the situation in problems with material numbers looks. Insignificant loss, and that not in all variants of optimization of a code, is observed only in problems 168.wupwise, 177.mesa, 188.ammp and 200.sixtrack. Other problems show a clear advantage on Core Duo, down to phenomenal sizes of the order of 1 % (!) In a problem 171.swim. Accordingly, integrated estimation rugged notebook review SPECfp_base20 on Core Duo T25 also appears on 29- % of more advantageous in comparison with more high-frequency Pentium M 770. Thus, productivity of kernel Yonah in problems with material numbers appears noticeably above in comparison with similar (on clock frequency and volume of a L2-cache) processor kernel Dothan. The similar result, but not with so high numerical parameters, was observed and above, by comparison of processor kernels Presler and Prescott. As processor Core Duo also is dvuh`yaderny'm, we shall estimate at last efficiency of parallel start of two spears of tests SPEC CPU20 in the metrics rate, having compared with results with those for rate-start of one copy of tests.
In integer problems outstanding again is 181.mcf, showing in this case not only absence of a gain, but rugged notebook review also some decline of productivity at start of two spears. The reason of this phenomenon to understand easily enough it is obvious, this problem not only is exacting to PSP, but also operates rather with great volumes of data so the greater size of a L2-cache promotes increase in its productivity. And as at two-line start each of kernels as a first approximation receives only half of total amount of a L2-cache, decrease in its productivity is observed. Despite of reduction of volume of a L2-cache in recalculation on a kernel, all other problems receive osch'utimoe advantage at transition to dvuh`yadernomu to a mode of the test. The least advantage is observed in a problem 300.twolf and makes 36- %, the greatest in a problem 252.eon and comes nearer to a theoretical maximum in rugged notebook review 1 %. By integrated estimation SPECint_rate2000, the gain in speed in dvuh`yadernom a mode on the average makes 67- %.
Problems with material numbers as a whole show not less successful picture. Except for 179.art for which parallel start harms more likely, than helps to business (we shall remind, the same was observed and on processor kernel Presler), problems with material numbers show a quite good gain in the speed, in some cases rather close coming nearer to a theoretical maximum. Nevertheless, by integrated estimation SPECfp_rate20 the prize as a whole in this class of problems appears a little bit smaller and makes 48- %. The conclusion the Investigated 65-nanometers processor kernels Presler (on an example of processor Pentium Extreme Edition 965) and Yonah rugged notebook review (on an example of processor Core Duo T2500) in integer tests SPEC CPU20 have shown the result at least little bit surpassing those for the previous generations of processor kernels Prescott and Dothan, accordingly, under condition of equality of volume of a L2-cache and some odds on clock frequency in favour of 90-nanometer of kernels. In tests SPEC CPU20 with material numbers processor kernel Presler shows a little bit greater in comparison with integer problems (up to % in SPECfp_base2000), and kernel Yonah and at all significant (up to % in SPECfp_base2000) advantage in comparison with processor kernels Prescott and Dothan, accordingly. Researches of efficiency of parallel start of problems in rugged notebook review the metrics rate have shown, that the majority of test problems SPEC CPU2000, not exacting to volume of a L2-cache and throughput of memory, are capable to show osch'utimy'y a gain in productivity at start of several spears of the appendix equal to quantity of system processors ( in case of Pentium Extreme Edition 965, in case of Core Duo T2500). Owing to additional use of technology Hyper-Threading, this gain makes 104-1 % for integer problems and 54- % for problems with material numbers on Presler whereas absolute parameters of a gain on kernel Yonah make 67- % in integer problems and 48- % in problems with material numbers.
Other discussions rugged notebook review in conference: Comments? Amendments? Additions?
The reprint of materials without the permission of edition is forbidden. For
the maintenance of advertisements the responsibility is born by advertizers.
