Benchmark testing is essential in evaluating performance gains or losses from overclocking. Two types of benchmarking techniques are available: synthetic and real world. Synthetic benchmarks use a series of performance-testing algorithms and enable you to compare your system’s results with results generated from other systems, across all platforms. Synthetic benchmark results for a single system are not useful; they must be compared to results obtained on other systems. Real-world benchmarks test actual performance for common operating scenarios, in real time.
Results are often measured in frames per second or operations per second.
Benchmark results can provide a foundation for determining the best combination of multiplier and bus overclocking techniques. Increasing system bus rates, while decreasing or maintaining processor multiplier settings, may produce better performance than raising multiplier values alone. A thorough benchmarking process can help you analyze the various impacts on performance and stability introduced by overclocking.
It is important to establish baseline performance numbers on your system before you do any overclocking, so that you can make before-and-after comparisons.
Overclocking does not always improve computing results; sometimes it brings negligible gains or even declines in performance. Before-and-after comparisons will help you determine whether or not your overclocking efforts have succeeded and if further adjustments are necessary.
System stability is often affected by overclocking, so it is valuable to determine what impact your efforts have had on stability without having to experience failures (that is, system hang-ups and crashes) in your real-world applications. Benchmarking apps involve strenuous system tests that exceed the level of CPU power and resources required by typical applications in everyday use. During such tests, a system can be pushed to its limits. If the system fails, it is probably either overclocked beyond its capabilities or in need of additional tweaks. Increasing CPU voltage, improving the cooling, or reducing bus speed while increasing the CPU multiplier may yield better results.
A few simple steps will ensure that your system is properly configured for benchmarking.
1. Your desktop display resolution should always be set to the same value for each pass or test run of a given application. A resolution of 1024 x 768 pixels with 16-bit color is recommended for consistency and comparable results.
2. All hard disks in your system should be thoroughly defragmented before testing to ensure maximum bandwidth, low access latencies, and a consistent data flow among each of the system components and the drive array.
3. Finally, all components (video, chipset, etc.) should be installed using the latest software drivers from each hardware manufacturer.
SiSoft Sandra is one of the most widely used synthetic benchmarking applications. Sandra offers performance testing for many subsystems and components. It also provides detailed analyses of system performance, capabilities, and stability.
Overclockers will appreciate Sandra’s dedicated tests designed for specific areas, such as processor, memory, and drive systems. These tests will help determine the best combination of settings for memory bus rates, front-side bus rates, and processor multiplier values.
Sandra’s processor testing is broken down into two benchmarking modules. The primary processor-testing module provides instructions-per-second ratings for both integer and floating-point math operations. The multimedia module tests any enhanced streaming instruction capabilities the processor may offer, including Intel’s SSE2 or AMD’s 3DNow!. While synthetic processor testing may not represent real- world performance, it is a good place to start in identifying changes that result from overclocking.
Figure 9-1: Sandra processor benchmark
The Sandra memory module lets you benchmark the available bandwidth between the memory and processor buses via the chipset bus. The results are displayed in megabytes per second for both integer and floating-point operations. Increases in front-side bus and memory bus rates offer greater bandwidth improvements than multiplier-only overclocking.
The Sandra file system benchmark provides both stability and performance testing. Front-side bus overclocking can produce higher drive bandwidth, with less read/write latency, through consequent overclocking of the PCI bus transfer rate. Stability can be evaluated by way of multiple-loop drive testing (i.e., running the test multiple times) to verify that data integrity is maintained during operation at extended PCI bus frequencies.
Figure 9-2: Sandra memory bandwidth benchmark
Figure 9-3: Sandra file system benchmark
MadOnion is the publisher of several popular synthetic benchmark applications. Its 3DMark series analyzes system performance at both the processor and subsystem
levels. Though it was originally designed to evaluate 3D video performance, the stress 3DMark places on the processor, chipset, and memory buses allows 3DMark to effectively benchmark of overall system performance.
3DMark2000 can test the effectiveness of the processor alone. It separates the video card’s advanced 3D functions from the rendering pipeline in order to isolate the processor’s performance independent of the video card. Be sure to disable any hardware transform-and-lighting (T&L) operations when configuring the test environment, so that the processor-related tests rely totally on software rendering routines. The 3DMark2000 also offers a looping demo mode that can be used to test system stability over time. In the SE version, internal rendering pipelines have been updated to support Microsoft’s DirectX 8.1 D3D hardware acceleration routines.
Figure 9-4: 3DMark2001 testing
Ziff Davis WinBench 99
Ziff Davis, the popular content publishing and hardware testing corporation, provides WinBench 99, version 2.0, as a free download. All WinBench testing routines are performed through 32-bit operations, with benchmarks designed to stress and evaluate graphics, video, and disk performance. The numbers generated by Graphics WinMark tests are based on comparative normalized scores, while the Disk WinMark scores represent actual drive transfer rates in thousands of bytes per second. Test results can be saved for later system-to-system comparisons or to analyze performance changes induced by various overclocking settings compared to your system’s baseline performance.
Figure 9-5: WinBench 99 comparison tables
Real-world Testing: 3D Games
Complex 3D games can test real-world performance by measuring rendering rates in frames per second. ID Software’s Quake 3 remains the dominant choice for such testing, even though its rendering engine is aging. Other popular games, like Unreal Tournament, Aquanox, and any other game with a frames-per-second benchmark capability, should prove equally acceptable. DirectX and OpenGL games lack internal testing functions, but they can still be benchmarked using the FRAPS performance measurement utility (http://www.fraps.com).
Figure 9-6: Quake 3 time demo results
Quake 3 frames-per-second testing is performed via the console command line interface.
1. Press the tilde (~) key while the game is running. This will invoke the interface.
2. At the command line, type timedemo 1 and press the ENTER key.
3. Close the console interface by pressing the tilde key (~) a second time.
4. Perform frames-per-second testing by selecting either available demo via the main menu. The average rates will be returned to the console interface when the demo is completed.
5. Record the values for later analysis.
6. Return the game to the original timedemo 0 state via the console interface to re-establish normal play.
Real-world Testing: Applications
Many multimedia applications can be used for real-world performance testing. Encoding sample video clips with Microsoft Windows Media Encoder, VirtualDub, or Adobe Premiere can be a very effective comparative test, as long as each encoding pass is accomplished under the same system configuration. Each of these applications can return values that indicate the amount of time elapsed during encoding. Similar operations can also measure the time required to apply transforms or filters to graphics in applications such as GIMP or Adobe Photoshop.
Figure 9-7: VirtualDub video compression