Core 2 Duo Scaling

February 05, 2007 at 12:18:00 AM, by Blair Mathis

Determine which Intel Core 2 Duo processor model is right for you!

Laptop Scaling

Intel's Core 2 Duo is easily the best mobile processor on the market today, boasting sky-high performance and low power consumption. But which CPU is for you: the super expensive T7600, the bargain-priced T5500, or something in between?

This article will first take a look at the various versions of Core 2 Duo that Intel has to offer. We will compare Core 2 Duo clocked at four speed grades: 2.33GHz, 2.16GHz, 2GHz, and 1.83GHz in a variety of benchmarks and let you know how to get your best bang for the buck.

Note: We will refer to a Core 2 Duo T7000' multiple times in this article. We were unable to get a Merom sample with 2MB L2 cache, thus we lowered the clock speed of our 4MB L2 model to 1.83GHz to perform near that of the T5600 (1.83GHz, 2MB L2 cache), resulting in our unofficial T7000' nomenclature.

The Core 2 Duo Processor

Intel has two lines of Core 2 Duo processors: notebook and desktop. Core 2 Duo for notebooks is codenamed Merom, while Core 2 Duo for desktops is codenamed Conroe. Merom is essentially a lower-clocked version of the desktop Core 2 Duo, codenamed Conroe. Like Conroe, Merom comes in both 4MB and 2MB shared L2 cache versions. Overall, Merom is clocked lower than Conroe, although there is some overlap. Conroe clock speeds range from 1.86GHz to 2.93GHz, while Merom clock speeds range from 1.66GHz to 2.33GHz. All Merom processors have a lower bus speed than Conroe - 667MHz versus 1066MHz - but are rated at half the Thermal Design Power: 35W TDP versus 68W TDP.

Model	Frequency	L2 Cache	FSB
T7600	2.33 GHz	4MB shared	667MHz
T7400	2.16 GHz	4MB shared	667MHz
T7200	2.00 GHz	4MB shared	667 MHz
T5600	1.83 GHz	2MB shared	667 MHz
T5500	1.66 GHz	2MB shared	667 MHz

The models with 4MB L2 caches have 16-way set associative L2 caches while the 2MB L2 caches are 8-way set associative. Higher associativity means that the cache has more possible locations to put data and is independent of the cache size. In other words, higher associativity means a more efficient cache. These findings were made with CPU-Z's Latency utility.

Benchmark Setup

For the tests, we were graciously provided with an Asus Z96J laptop by Intel. The laptop was set up with the following basic configuration:

Component	Asus Z96J
CPU	Intel Core 2 Duo T7600
RAM	1x1GB PC2-5300 (667MHz) CAS5
Graphics	ATI Mobility Radeon X1600 256MB
Hard Drive	80GB 5400RPM 8MB cache

To run the processor at different clock speeds, we used RM Clock to change the CPU multiplier. We tested the T7600 between 2.33 GHz and 1.83 GHz. Please note that since the T7600 has a 4MB L2 cache, the 1.83 GHz datapoint is merely for interest's sake only. Models with 2MB L2 cache are slightly slower than models with 4MB L2 cache.

Synthetic

Synthetic benchmarks attempt to determine performance by testing various characteristics of the computer. Since synthetic benchmarks are not always representative of true CPU performance, very few CPU-centric synthetic benchmarks were tested.

PCMark 2005

PCMark 2005 1.0.1 is the latest update to Futuremark's popular overall system benchmarking program. The 2005 version adds multithreading, DirectX 9, Windows Media Player 10, virus scanning, High Definition video playback (WMVHD), and a vast number of other tests to its suite. Testing your computer's CPU, RAM, hard drive and graphics card, PCMark05 drives your computer to the max to determine its strengths and weaknesses.

Benchmark	2.33 GHz	2.16 GHz	2.00 GHz	1.83 GHz
CPU	5945	5511	5088	4683
RAM	4514	4308	4044	3841
GFX	3155	3131	3109	3083
HDD	3966	3955	3877	3928

The CPU subscore for PCMark 2005 increase by roughly 8% every 166 MHz. Somewhat unexpectedly, the RAM subscore also increased by 4-5% for every 166 MHz step. PCMark's RAM subtests are clearly slightly CPU limited, showing corresponding improvements in the RAM read/write/copy/latency tests. Also, the graphics score increased by roughly 1% every 166 MHz, showing the improved data processing efficiency.

Increasing CPU frequency only improves performance around 4.3% per 166MHz step for the overall score. This makes sense since PCMark 2005 tests your entire system - CPU, RAM, GPU, and HDD - and CPU performance is only one part of the system's overall performance.

Gaming

Another major motivation for CPU performance is in the realm of gaming. We tried benchmarking Quake 4 at a minimum resolution of 800x600 and quality setting of Medium. However, CPU speed did not affect performance at all, even when turning on multithreading in-game, remaining at a steady 29.95 FPS. It appears that Quake 4 is completely GPU-limited due to the test machine's Radeon X1600 256MB graphics card.

Therefore, since 3DMark scores showed a difference - albeit a very small one - we only have 3DMark scores available. Treat the 3DMark scores with caution: the scores most likely do not reflect in-game FPS scores.

3DMark 2003

3DMark 2003 shows negligible increases in score - less than 1% - with increasing frequency. Since 3DMark 2003 is fairly representative of an older game at 1024x768 resolution, we can see that even older games are GPU limited in a notebook like this.

3DMark 2006

Since the 3DMark composite scores were uninteresting (scores ranged from 1740 to 1771), only the CPU score will be shown here.

The 3DMark 2006 CPU benchmark tests your computer's ability to perform AI functions, such as pathfinding. By creating a CPU-limited situation, we see a large increase in performance with clock speed at roughly 8% with every 166MHz step.

Scientific

Improving scientific and engineering-related calculation performance is one of the main driving factors of computer development. Since many such tasks can take days to complete, scientists and engineers are always highly interested in the latest CPU innovations.

Here we will take a look at several scientific/engineering benchmark applications: SuperPI, ScienceMark, Mathematica, and SPECViewPerf 9.0.

SuperPI
SuperPI calculates the irrational number Pi up to a certain number of digits. In August 1995, the calculation of Pi up to 4,294,960,000 decimal digits, the world record at the time, was achieved by using a supercomputer at the University of Tokyo. The program was written by D. Takahashi with collaboration from Dr. Y. Kanada at the University of Tokyo. This record-breaking program was ported to personal computer environments such as Windows NT and Windows 95 and called SuperPI.

1M
We first calculated Pi to 1 million digits. This benchmark is small and can fit entirely in Core 2 Duo's large 4MB shared L2 cache.

Increasing CPU frequency improves performance greatly, at 8-10% per 166 MHz division.

32M
However, when we increase the benchmark to 32 million digits, the benchmark no longer fits in Core 2 Duo's L2 cache. Most of the benchmark is stored in system memory. Therefore, this benchmark is less dependent on CPU speed and more dependent on RAM speed.

Since this benchmark also depends heavily on RAM bandwidth, we can start to see the speedup decrease more and more with increasing clock speed. Speedup ranges from 4% to 6% per 166 MHz division.

Sciencemark
Sciencemark 2.0 is a collection of scientifically-based benchmarks, performing real-world calculations. In an attempt to model real world demands and performance, SM2 is a suite of high-performance benchmarks that realistically stress system performance without architectural bias.

Cypher
Cypher is an encryption benchmark, measuring performance in the popular AES 128/256 and RSA 512/1024 algorithms.

Cypher also shows relatively strong gains with increasing CPU speed, improving by 7-9% per 166 MHz division.

MolDyn
Molecular Dynamics is a method for simulating the thermodynamic behavior of materials using their forces, velocities, and positions.

MolDyn shows 7-8% improvement per 166MHz division.

Primordia
Primordia, meaning "atoms" in Latin, computes the electric orbitals for the atoms between Hydrogen and Promethium in the periodic table using a Restricted Hartree-Fock method. Primordia outputs the total all electron energy of the atom in addition to the kinetic and potential contributions. The user is allowed to specify different grids upon which the orbitals are determined. Default settings were used.

Increasing CPU frequency improves Primordia scores by 6-8% per 166MHz.

Mathematica 5.2

Mathematica is a popular mathematics program for the academic and engineering industry, combining many powerful computational tools into a single package. Rather convenient for us, Mathematica also has a built-in benchmark system, which can be loaded with the Benchmark utility. Mathematica performs a number of commonly used computations, such as calculating Pi, Fourier transforms, and solving linear systems.

Each individual benchmark score is in the table below. All times are in seconds.

Benchmark	2.33 GHz	2.16 GHz	2.00 GHz	1.83 GHz
Data Fitting	1.42	1.59	1.66	1.86
Digits of PI	1.77	1.88	2.09	2.27
Discrete Fourier Transform	1.64	1.72	1.88	1.92
Eigenvalues of a Matrix	4.25	4.63	4.98	5.42
Elementary Functions	4.67	4.96	5.53	5.97
Gamma Function	1.77	1.91	2.06	2.25
Large Integer Multiplication	1.88	1.98	2.17	2.38
Matrix Exponential	2.58	2.80	3.05	3.30
Matrix Multiplication	14.6	15.7	17.1	18.6
Matrix Transpose	2.66	2.86	3.02	3.31
Numerical Integration	2.34	2.50	2.75	3.00
Polynomial Expansion	1.59	1.66	1.95	2.16
Random Number Sort	1.98	2.11	2.30	2.45
Singular Value Decomposition	6.27	6.66	7.25	7.88
Solving a Linear System	9.72	10.41	11.40	12.1
Total Time	59.1	63.3	69.3	74.9

Mathematica then takes the results and calculates a total score.

Increasing CPU speed improves benchmark results by 7-10% per 166MHz stepping.

SPECViewPerf 9.0
SPECViewPerf 9.0 is an OpenGL graphics performance benchmark which tests rendering capability in a variety of popular programs: 3ds max, CATIA, EnSight, Lightscape, Maya, Pro/ENGINEER, Solidworks, and Unigraphics.

Performance is measured in a weighted geometric mean of units FPS (frames per second).

Benchmark	2.33 GHz	2.16 GHz	2.00 GHz	1.83 GHz	Avg Speedup
3dsmax-04	9.375	9.134	8.647	8.388	3.8%
catia-02	11.32	10.95	10.27	9.873	4.7%
ensight-03	8.892	8.728	8.394	8.280	2.4%
light-08	11.00	10.34	9.386	8.826	7.6%
maya-02	13.00	12.44	11.65	11.17	5.1%
proe-04	6.634	6.466	6.218	6.136	2.6%
sw-01	12.93	12.77	12.50	12.46	1.2%
ugnx-01	6.553	6.492	6.346	6.310	1.3%
tcvis-01	2.942	2.944	2.937	2.927	0.2%

As we can see, we have a mixed bag in terms of CPU scaling with performance. Programs such as EnSight, Pro/ENGINEER, SolidWorks, and Unigraphics exhibit small performance gains with increasing CPU speed, while programs like 3ds max, CATIA, Lightscape, and Maya exhibit much more noticible performance gains. Of particular interest is light-08, which is a trace of Discreet's Lightscape radiosity program, which gains an amazing 7.6% average increase in speedup per 166 MHz step. Although SPECViewPerf is defined to be a graphics performance benchmark, it is much more CPU-intensive than your average game.

Multimedia

The most common heavily CPU-limited tasks performed would have to belong in the multimedia category. Video and audio encoding, among other jobs, can take upwards of an hour to complete. Therefore, for people who frequently perform these tasks, it is of crucial importance to purchase a computer with a powerful CPU.

Cinebench
Cinebench is the free benchmarking tool for Windows and Mac OS based on the powerful 3D software CINEMA 4D. The tool is set to deliver accurate benchmarks by testing not only a computer's raw processing speed but also all other areas that affect system performance such as OpenGL, multithreading, and multiprocessors.
Cinebench includes render tasks that test the performance of up to 16 multiprocessors on the same computer as well as software-only shading tests and OpenGL shading tests on huge numbers of animated polygons that will push any computer to its limits.

Cinebench, like most multimedia applications, scales wonderfully with increasing CPU frequency, improving roughly 8% every 166MHz.

XMPEG 5.0.2
XMPEG is a popular video encoding software frequently chosen for its ease in use. We encoded a short video file (ATI-X800-Subsurface-Demo-v1.0) from MPG format to DivX at a bitrate of 226Kbps.

Increasing clock speed in XMPEG (as well as other encoding software) shows nearly linear speed increases of around 7.3% to 8.7% per 166MHz division.

LAME MT

For the MP3 encoding benchmark, we chose a multithreaded version of the LAME MP3 encoding program. LAME MT is based off LAME 3.97 Alpha.

The website provides a simple benchmark to test performance. Interestingly enough, it includes both Intel and Microsoft compiled versions of the code. Even more interesting is the fact that the Intel version of the code is much faster than the Microsoft version of the code.

Benchmark	2.33 GHz	2.16 GHz	2.00 GHz	1.83 GHz
Intel, no multithreading	34	37	40	44
MS, no multithreading	42	46	50	55
Intel, multithreading	20	22	24	30
MS, multithreading	29	25	27	26

Increasing clock speed yields tremendous speedup of 8-9% per 166 MHz stepping.

Photoshop CS2
Adobe Photoshop CS2 is one of the most popular and powerful image editing software on the market today. To benchmark this program, we used DriverHeaven's Photoshop Benchmark V2, which tests a number of filters and actions with a huge image file.

Benchmark	2.33 GHz	2.16 GHz	2.00 GHz	1.83 GHz
Texturizer	1.2	1.3	1.4	1.4
CMYK Color Conversion	2.1	2.3	2.4	2.6
RGB Color Conversion	2.4	2.7	2.8	2.9
Dust and Scratches	8.1	8.6	7.8	9.3
Watercolor	15.4	15.7	16.9	18.3
Texturizer	7.4	8.2	7.0	7.2
Stained Glass	6.7	6.9	6.6	7.8
Lighting Effects	9.4	9.1	9.5	9.5
Mosaic Tiles	13.7	13.2	13.9	15.3
Extrude	48.4	50.9	54.5	59.0
Smart Blur	44.1	45.0	48.5	52.5
Underpainting	15.3	16.4	17.8	19.6
Total	174.2	180.3	189.1	205.4

Since Photoshop is much more RAM bandwidth limited than CPU limited, we only see small speedup increases of around 3-5%. The exception is the jump from 1.83 GHz to 2.0 GHz, which sees a speedup of an amazing 8%.

WinRAR
WinRAR is a popular compression tool for Windows XP with a built-in benchmark system. We ran the benchmark until the result stabilized.

Increasing clock speed in WinRAR nets slow but steady gains of around 2.5% to 3% per 166MHz.

Analysis and Conclusion

As shown from the benchmarks, every 166 MHz increase in clock speed yields a 0-10% increase in performance. Generally, for CPU-limited applications, performance increases ranged around 6-8% per 166 MHz step. So now the question becomes, are the higher clock speed models worth it?

To answer this question, we took a look at the current pricing for the Core 2 Duo T7200, T7400, and T7600 models. Since the T7000 does not exist, we also looked at the T5600 (keep in mind that our performance data does not reflect the performance of the T5600 due to L2 cache differences).We compared pricing not only for individual CPUs from the popular online store Newegg, but also for CPU upgrades from Dell. To calculate the cost for the CPU upgrade, we took the baseline price for a Dell Latitude D620 configured with the T7200 and calculated the price difference for the same model with a T7400 or a T7600. We did the same thing for a Lenovo ThinkPad T60 and an HP nw8430 (all three are business-class machines).