The following benchmarks have been compiled from PC Magazine's CPU Article, which can be found at ZDNet's website.
ZD Business Winstone 97
The ZD Business Winstone 97 test produces a number that reflects a tested system's overall performance in a real business setting. In that sense, Winstone 97 is a better measure of a CPU's true effectiveness than even CPUmark.
The Pentium II processors scored best under all memory/OS configurations, but because of heavy use of I/O in running the applications that make up the Business Winstone test suite, sheer processor muscle does not shine through. As a result, there isn't much performance variation among the processors. Under the Windows NT/64MB configuration, the difference between the best CPU, the Pentium II/300, and the worst, the Pentium/200, is only 48 percent--a meager improvement considering the price difference. The Cyrix 6x86MX-PR233 edged AMD's K6/233 and almost matched the Pentium II/233 on all Business Winstone runs. The Cyrix 6x86-PR200 outperformed (slightly) its Intel counterpart, the 200-MHz classic Pentium, and surprisingly, so did the AMD K5-PR166. All CPUs showed about the same relative performance under both operating systems, except the Pentium Pro/200. Intel optimized this sixth-generation CPU to run true 32-bit operating systems like Windows NT. True to its calling, the Pentium Pro was in the 233-MHz K6 and 6x86MX class under Windows NT, but was barely faster than a classic Pentium/200 under Windows 95.
There's minimal gain in upgrading to 64MB from 32MB under Windows 95. But under Windows NT, where the OS occupies a larger memory footprint, the investment in extra memory brings a significant return of up to 18 percent in improved performance. Also, it was found that the faster processors take better advantage of extra memory; as the hard disk's involvement is reduced, the CPU's power becomes more evident.
To empirically demonstrate the effect of a faster hard disk on the reduced I/O bottleneck, Business Winstone 97 was run with 64MB under Windows NT with a 10,000-rpm Ultra SCSI disk instead of our standard 5,400-rpm EIDE disk. This provided a minimal 1 to 2 percent benefit with most CPUs, while the fastest CPUs had improvements of 4 to 5 percent--better, but not significantly so. In scaling up Pentium II speeds, there was minimal gain in going from 266 to 300 MHz, mainly due to the inadequacy of current system bus and I/O technologies.
ZD High-End Winstone 97
The computation-intensive applications in High-End Winstone 97 emphasize a CPU's abilities and include more floating-point calculations than Business Winstone. Because this test also involves applications with bigger memory footprints, it not only isolates the CPU's performance better than does Business Winstone, it also shows more drastic memory-scaling effects.
The Pentium II distanced itself from the field in this test. For example, under Windows NT with 64MB, the Pentium II/233 outperformed the AMD K6/233 and the Cyrix 6x86MX-PR233 by a modest 8 and 6 percent, respectively, on Business Winstone, but the gap widened to 26 and 29 percent, respectively, on High-End Winstone. This change is also partially attributable to the Pentium II's superior floating-point ability. Whereas the Pentium II gained ground, AMD and Cyrix lost it: The small performance edge that the AMD K6 and the Cyrix 6x86MX held over the Pentium MMX/233 on Business Winstone shrank on High-End Winstone.
The Pentium Pro/200, on the other hand, excelled on High-End Winstone, proving itself a superior workstation processor to the K6 and the 6x86MX. From these findings it can be deduced that Intel's architecture is more proficient in processing high-end applications than AMD's and Cyrix's, which are optimized for business software. It is also quite possible that the programmers of processor-intensive applications tune their codes for Intel CPUs.
More memory has a bigger impact on High-End Winstone results than on Business Winstone results. In upgrading from 32MB to 64MB, there were improvements ranging between 10 and 18 percent on Business Winstone under Windows NT. On High-End Winstone, the performance boost ranged from 24 to 54 percent. Again, faster CPUs take better advantage of additional memory.
3D WinMark 97
As faster 3-D accelerators relieve the CPU of more rendering tasks, they should reduce the CPU's contribution to 3D WinMark. True to this theory, the variance in scores attenuated as we went from no accelerator hardware to an average 3-D card (the Diamond Stealth 3D 2000XL using an S3 ViRGE chip) to an above-average 3-D card (Leadtek WinFast 3D L2200 with 3Dlabs' Permedia NT chip set, which includes the Glint Delta chip).
More interestingly, we discovered that the 3-D performance of faster CPUs such as the Pentium II is actually slowed when they have to delegate 3-D rendering to average accelerators such as the Diamond 3D card. To verify the 3-D "deceleration" effect, we isolated the scores from scene 1, which is rendered completely in hardware, for both 3-D cards. Indeed, the Pentium II was faster when rendering in the emulation mode than with the Diamond card. Therefore, we recommend using only superior 3-D cards like Leadtek's if you have a Pentium II-based system. All other CPUs showed benefits when used with either of the 3-D accelerators. The test was not performed with dual CPUs for one of the tests, because that benchmark test was not multithreaded and would not have taken advantage of the second CPU.
To test floating-point performance, which affects high-precision scientific calculations or complex graphics rendering, a synthetic floating-point test was used, which is not affected by memory size.
The two tests agreed except on the K6's performance relative to other CPUs. The K6 fared worse in the synthetic test compared to the classic Pentium; the K6's internal architecture may allow it to perform better in memory-intensive operations compared with the classic Pentium. Intel's pipelined floating-point architecture proved superior to both AMD's and Cyrix's in all other cases. The Pentium/200 outscored the 6x86MX-PR233 on both tests and the Pentium II/233's score on the synthetic test was 72 percent better than that of the K6/233. The Pentium Pro also showed strong floating-point performance, outclassed only by that of the Pentium II.
To aquire these benchmarks, machines with identical subsystem hardware were used, in order to isolate CPU performance. Each unit had EDO memory, an IBM 5GB EIDE hard disk, and a Diamond Stealth 3D 2000XL graphics adapter. The L2 cache was set to 512K, except for the Pentium Pro, which has 256K of L2 cache embedded on the chip. The testing was performed under two operating systems, Microsoft Windows 95 with OSR 2 and Windows NT 4.0 with SP 2. A common motherboard with all the CPUs, however, was not used, because no single motherboard supports all the processors we tested and because motherboards are often tuned to specific CPUs before being sold. All tests were performed with the display set to 1,024-by-768 resolution with 65,536 colors, except for ZD High-End Winstone 97, AutoCAD, and Pro/Engineer, which were ran at 1,024-by-768 with 16.7 million colors.
ZD Winstone 97, Version 1.0, benchmark test suite was used to measure the machines' performance when running 32-bit Windows applications. Winstone 97 is divided into two categories: ZD Business Winstone 97 and ZD High-End Winstone 97. Business Winstone measures the time a PC takes to execute a set of application scripts that use eight of the best-selling Windows 95 applications. Business Winstone weights a given PC's test timings based on the applications' unit market shares; it derives a composite number and converts it to a relative score. High-End Winstone 97 measures the time it takes for a system to execute a set of application scripts that exercise six 32-bit Windows applications in four categories: CAD, application development, image editing, and 3-D visualization. High-End Winstone weights a given machine's performance on the six tests equally, then derives a composite number and converts it to a relative score. (The score is relative to the performance of a Dell Pentium/100 system with 32MB of RAM, whose score is defined as 10.)
To measure the MMX performance, we set the testing parameters to engage the MMX DLL and force all 3-D rendering to be executed by the CPU using Direct3D's emulation library. To study the CPU's relation to 3-D accelerator hardware, we tested with two 3-D cards: an average one (Diamond Stealth 3D 2000XL using the S3 ViRGE chip set) and an above-average model (Leadtek WinFast 3D L2200 using 3Dlabs Permedia NT). We ran the 3D WinMark suite of tests only under Windows 95 with 32MB of memory, a common configuration for Direct3D games.
Thanks to PC Magazine and Ziff-Davis Benchmarks for providing the Benchmark Programs, and scores.
If you would like to compare your CPU to the scores found above, you can download ZD's benchmarking software from http://www.zdbop.com/.