Contents

Summary

Tests were run on a new PC, with a 3 GHz Phenom II processor, a GeForce GTS 250 graphics card and Widows 7, to measure system temperatures and multiprocessor performance. Temperature measurement available were CPU case, CPU core, motherboard and graphics processor. A slightly up market copper heatpipe CPU cooler was fitted, and this might explain why the CPU case temperature, normally at room temperature + 13°C when idling, only increased by a maximum of 14°C. Initially, the processor ran at 800 MHz and the first 2°C rise was almost instantaneous with the CPU switching to full speed. Maximum temperature was often reached within 5 minutes. Reported CPU core temperatures were lower than those for the case but increases under load were higher. It then became apparent that the reported core temperatures are not real but used purely for thermal management purposes, with a maximum value of 70°C.

IntBurn64 - this carries out 64 bit integer arithmetic with high speed data from caches or RAM. Four different sizes were used for data in dedicated L1 and L2 caches, shared L3 cache and RAM, to use one then four processors. With dedicated caches, average throughput gains in using four CPUs was 3.94. The L1 cache test produced the highest speeds at 17,914 Millions of Instructions Per Second (MIPS) but maximum case/core temperature increases of 14/18°C were produced by the L2 cache test, at up to 12,613 MIPS. For the L3 cache tests, each processor requested 1 MB of the 6 MB shared space. Resulting gains on using four CPUs averaged 3.5 times. Temperatures were similar to those for L1 cache tests. RAM speed with one CPU was less than a third of possible maximum data transfer rate but achieved 68%, or 14.5 GB/second, using four processors.

SSEBurn64 - this uses SSE or SSE2 floating point instructions and has CPU only and L1 cache tests, with resultant speeds measured in Millions of FLOating Point instructions per Second (MFLOPS). Maximum 32 bit SSE MFLOPS for a 3 GHz processor is 12,000, or possibly 24,000 with linked add and multiply instructions. Maximum 64 bit SSE2 MFLOPS are half those for SSE. Memory size has to be specified for other tests that stress caches or RAM, with results calculated as MB/second. Performance gains, using four processors, were similar to the integer tests. The SSE CPU test, using registers instead of cache, ran at nearly 12,000 MFLOPS on each CPU but temperature increases were the lowest. The SSE Cache test generated 16,400 MFLOPS per CPU and produced the same high temperatures as those with the integer tests. Memory Tests - L2 cache tests gave rise to the highest temperatures of this group, with a data transfer speed half that using L1 cache.

Livermore Loops - This was the original key benchmark for supercomputers, having 24 kernels of numerical applications with speeds calculated in MFLOPS. The benchmark was known to produced errors on an earlier overclocked processor. Here, maximum temperatures were moderately high, with four CPU each running at up to nearly 3900 MFLOPS (i387 instructions not SSE type).

VideoD3D9 - This is a DirectX 9 benchmark where any one of the 8 tests can be run as a burn-in program at a specified window size. Speeds are recorded in Frames Per Second (FPS). The tests were run individually to find the hottest one. This was then run at the same time as three SSE Cache tests. Recorded temperature increases were just about the highest - Case +14 to 48°C, Core +19 to 45°C, Board +3 to 36°C and Graphic Processor +26 to 69°C. SSE MFLOPS were degraded somewhat, as the graphics test uses the equivalent of 120% CPU utilisation.

Other Tests - The graphics/SSE test was repeated using integrated Radeon HD 4200 graphics on the Asus motherboard. The results were much lower temperatures and FPS but higher MFLOPS. CUDA provides nVidia programming extensions to use a GPU for general purpose computing. The test using these ran at 159,600 MFLOPS with temperature of the GPU changing by +30 to 73°C. OpenMP functions enable shared data calculations over available CPUs. The MS compiler is not efficient in translating to SSE instructions, only achieving 3700 MFLOPS on one CPU and 14,446 MFLOPS using four, with associated low temperature increases. A game benchmark was run a number of times and this increased GPU temperature by 29°C. It used more than three CPUs but heating effects were not high.

Reliability or Burn-in Benchmarks

This report provides results of a series of performance and reliability tests on a 3.0 GHz Phenom II X4 CPU using 64-Bit Windows 7. The hardware comprises a 4 processor Phenom II 945 on an Asus M4A785TD-V motherboard, with 8 GB DCDDR3 RAM, a WD 5400 RPM Green SATA disk and a GeForce GTS 250 graphics card, plus Radeon HD 4200 on the motherboard. The processor has a Titan TTC-NK34TZ "Super Quiet 22dBA Triple Copper heatpipe CPU Cooler".

Reliability/Burn-in programs used were mainly compiled for 64 bit working, including tests using integer instructions, DirectX 3D graphics functions and floating point arithmetic calculations. CPU, motherboard and graphics processor temperatures were recorded as each test was run. Besides being used for stress testing, the programs provide useful performance information of multi-core processors. These results are saved in text log files, examples being given below. The compiled programs, source code, descriptions and performance results on numerous systems can be obtained via the links given below.

System Information

All my latest benchmarks and test programs include the following, where Windows NT Version 6.1, build 7600 indicates Windows 7. Memory shown could be 8192 MB but, with integrated graphics enabled, this is reduced by 256 MB.

Temperature Monitor

During the tests, CPU case and core temperatures were measured using CPUID Hardware Monitor and, when applicable, confirmed using Asus Probe II. The accuracy of these measurements is unknown and the temperatures and increases seem particularly low. Below is an example of maximum monitored temperatures using 4 CPUs, with room air at 21°C. The core temperatures are very low when the CPUs are idle, but suddenly increase significantly when the tests are started. Core voltage also increases from 0.93V to 1.3V. CPUZ indicated that that CPU was running at 800 MHz when idle, increasing to 3.0 GHz. Core/Case temperatures with the CPU running at 3 GHz, but idling, were typically around 4°C/ 2°C higher than at the lower MHz. All measurements produced maximum temperatures on the same sort of time scale and Prime95 showed similar maximum temperature.

Turning off Cool’n’Quiet in BIOS increases CPU idling speed to 3 GHz and, at the same room temperature, case and core temperatures to 36°C and 30°C respectively. These temperatures were also confirmed after installing AMD’s OverDrive utility. It was then found that there is only one Core measurement that is known as Tctl Processor Temperature Control Value. It does not represent an actual temperature but is a relative reading that can be used for thermal management. Maximum value is 70°C.

CPUID Hardware Monitor Value Min Max CPU VCore 0.94 V 0.93 V 1.30 V Temperatures CPU case 35 °C 34 °C 47 °C Motherboard 33 °C 33 °C 34 °C Fans CPU RPM 2732 2721 2755 Fan 2 RPM 1280 1273 1293 Temperatures Core #0 26 °C 25 °C 43 °C Core #1 26 °C 25 °C 43 °C Core #2 26 °C 25 °C 43 °C Core #3 26 °C 25 °C 43 °C

IntBurn64 Reliability Test

This program uses assembly code and eight 64 bit integer registers (r8 to r15) that are not used with 32 bit code. It has twelve tests, adding and subtracting different data patterns. The first six tests alternately write and read data and the others are read only. The results are checked for correct calculations. Drop down lists are provided to select memory size used, between 4 KB and RAM size minus 64 MB, to test using data in L1 cache, L2 cache or RAM. Another list allows running time to be selected between 1 and 1000 seconds per test. An example of the log file is shown below.

IntBurn64 MP Performance and Temperatures

The following show performance levels in MBytes per second and temperature increases, running four copies of IntBurn64 at the same time, plus the speed using one CPU. The total data transmission rates are calculated, performance gains using four processors and Millions of Instructions Per Second (MIPS) from known instruction counts. Optional data size parameters used were 4 KB, 256 KB, 1 MB and 64 MB. The tests were run for 5 minutes, the one producing the highest temperature gains repeated for 10 minutes.

SSEBurn64

SSEBurn64 uses SSE or SSE2 Single Instruction Multiple Data (SIMD) floating point instructions to soak test the CPU, Cache or RAM at high speeds whilst checking results for correct values. SSE and SSE2 Run buttons are provided for separate CPU, Cache and RAM tests. The program produces 1024 random floating point numbers used in all tests. For the CPU test, 32 add or multiply instructions manipulate a few at a time from registers within a loop. The Cache test uses the same 32 instructions but with data from L1 cache within the main loop. The RAM test is biased towards fast data transfer and can also use cache sized data. Every fifth pass the memory is filled with 16 or 32 of the random numbers with the first set being read and checked for correctness. The main loop uses 8 load/add and 8 load/subtract instructions to produce a sum check of zero. CPU and Cache tests check that results are the same as the first pass which also calibrates the testing loops to run for up to one second (on a fast CPU).

Livermore Loops

This was the original key benchmark for supercomputers, having 24 kernels of numerical applications with speeds calculated in Millions of Floating Point Operations Per Second or MFLOPS. Overall performance characteristics are identified by geometric, harmonic and arithmetic means, minimum and maximum. The program also checks the results for computational accuracy and this was adopted to check for consistent numeric results for a burn-in test.

VideoD3D9_64 and SSEBurn64

VideoD3D9 is a DirectX 9 benchmark where any one of the 8 tests can be run as a burn-in test at a specified window size. Speeds are logged in Frames Per Second (FPS) over each minute of the tests. With the CPU burn-in and Direct3D tests in the same folder, a BAT file, with the commands shown below, can run them both at the same time. Firstly, all DirectX 9 tests were run individually for 5 minutes at 1680 x 1050 pixel monitor setting to identify the hottest test. Perfmon performance monitor was run at the same time to log CPU utilisation, where that recorded is average per CPU.

Next, The Vertex Shader routine and three copies of SSEBurn64 cache tests were run for 10 minutes. The former produced 30% CPU utilisation or 120% of one CPU. This was reflected in multiprocessor performance, where MFLOPS per CPU was reduced from around 16,500 to 13,600. Maximum CPU case and motherboard temperatures were just about the highest at 48°C and 36°C. Asus Probe II monitor has default maximum settings of 77°C and 60°C but the AMD specification for the former appears to be 71°C. Maximum GeForce GTS 250 temperatures are shown as 105°C, 69°C being recorded for these tests.

Other Tests

CUDA, from nVidia, provides programming functions to use GeForce graphics processors for general purpose computing. These functions are easy to use in executing arithmetic instructions on numerous processing elements simultaneously. As a benchmark, tests are run using different data sizes and increasing numbers floating point calculations per data element, with and without transferring data from/to main processor RAM. The reliability test uses calculate only with 32 calculations per word.

OpenMP is a system independent set of procedures and software that arranges automatic parallel processing of shared memory data when more than one processor is provided. This option is available in the latest Microsoft C++ compilers. Potential performance gains due to hardware SIMD with SSE instructions are not realised due to compiler limitations and this enhances the comparative benefit of CUDA GPU parallel processing. The benchmark executes the same range of functions, using the same data sizes, as the CUDA benchmark, but only with data in and out.

OpenMP tests at least show that a quad processor can achieve up to a near four times performance gain over a single CPU, but the relative slow speed leads to low temperature increases. CPU temperatures are even lower with the CUDA test, with processor utilisation equivalent to 100% of one CPU. On the other hand, that huge MFLOPS speed gives rise to the highest graphics processor temperature. For further details of these programs, see references below.

Reference Files

SSEBurn64 and IntBurn64 - Benchmark and source code in More64Bit.zip - Further results and description in BurnIn64.htm

Livermore Loops - Benchmark and source code in Benchnt.zip - Results and description in Livermore Loops Results.htm

VideoD3D9_64 - Benchmark and source code in Video64.zip - Further results and description in 64 Bit Graphics Tests.htm and Direct3D Results.htm

CUDA - Benchmark and source code in CudaMFLOPS.zip - Further results and description in Cuda1.htm

OpenMP - Benchmark and source code in OpenMPMFLOPS.zip - Further results and description in OpenMP MFLOPS.htm

Other - Burnin32.htm, Burnin64.htm, Vista64.htm, Win64.htm