How to use Geekbench on Windows 10

From wiki.mikejung.biz
Jump to: navigation, search

How to use GeekBench3 and GeekBench4 on Windows 10[edit]

This page will cover how to run Geekbench3 and Geekbench4, specifically on Windows 10. If you are using Windows 7, or Windows 8 this guide should also apply to you.

Before Running GeekBench[edit]

Before running Geekbench4 you will want to make sure you close out any unneeded programs - even ones running in the background like Steam, Battle.net or whatever else you got running. While these background processes may not cause a high difference in the results it's always good to ensure you are testing in a somewhat if not exactly consistent enviroment.

GeekBench4 CPU and GPU Compute Benchmark Tests[edit]

GeekBench4 contains two main sections, which are CPU benchmarks and "Compute Benchmarks" which is really more like GPU compute, but whatever. Anyway, the sections below will cover what tests are run in each section and what ones are the most relevant.

For the most part you will want to use the CPU benchmarks to measure you CPU performance. If your CPU has integrated graphics (GPU), then you can use the compute benchmarks to get an idea of how that performs to a dedicated GPU.

If you want to measure your GPU compute performance, you can do so using the recently added compute benchmarks, which did not exist in GeekBench3. You can select from CUDA and OpenCL tests, which I'll cover later on.

Geekbench4-main-section-cpu-gpu-compute-benchmarks.jpg

CPU Tests - 64-bit[edit]

For the most part there is no reason to run a 32-bit test unless you are measuring a dinasaur aged machine, in which case measuring performnace seems like the least of your concerns ;) Geekbench 4 could complete all the single core and multi core tests within a matter of minutes, or a lot longer depending on how fast your CPU is. Once the tests complete running you will see an new window that contains single and multi core results.

At the top is the "GeekBench Score" for single and multi core performance. Underneath the general scores is the System Information section which is useful and worth keeping along with the test results (will show you how to export results later on).

Geekbench4-cpu-results-geekbench-score-system-info.jpg

You will want to pay attention to both the single and multi core scores. Depending on your use case / environment you may want to focus on one over the other (for example, in a cloud environment where not a vCPUs are equal).

single core - Useful for measuring Cloud vCPU performance. Higher single core scores indicate better single core performance. Apps that don't use multiple cores should run faster on CPUs that have higher single core scores.

multi core - Also useful for measuring Cloud vCPU performance, but also useful for measuring the performance of larger servers that have 16+ cores. Not every CPU scales as expected so be sure to pay attention to both single and multi performance.

GeekBench4 Single Core CPU Workloads[edit]

Geekbench4 runs many Single Core workloads that are used to create the final "GeekBench Single Core Score" that makes it easier to understand overall Single Core CPU performance. If you want to know specific workload performance, you will want to focus on the individual test results, which are shown below.

Geekbench4-cpu-results-single-core-detailed-workloads.jpg

Some notable workloads include

  • AES - Measured in GB/sec - If you care about encryption performance then you may be interested in workloads such as AES.
  • LZMA - Measured in MB/sec - Care about compression? workloads that use algorithms like LZMA should be focused on.
  • JPEG - Measured in Mpixels/sec - If you do photo editing or rendering, workloads like JPEG help determine photo performance.
  • Face Detection - Measured in Msubwindows/sec - If you are Mark Zuckerburg or the NSA, this is your test ;)
  • Memory Copy - Measured in GB/sec - Throughput test for Memory copying.
  • Memory Latency- Measured in Moperations/sec - Because this is a latency test, we are resuming memory operations per second, similar to iops, but in RAM.
  • Memory Bandwidth - Measured in GB/sec - Throughput test for raw RAM bandwidth.

GeekBench4 Single Core CPU Workloads[edit]

Geekbench4 also runs many Multi Core workloads that are used to create the final "GeekBench Multi Core Score" that makes it easier to understand overall CPU performance. If you want to know specific workload performance, you will want to focus on the individual test results, which are shown below. The workloads ran for Multi Core are the same as Single Core. This gives you useful metrics you can use for specific workloads to determine how well the application scales as core count rises.

Geekbench4-cpu-results-multi-core-detailed-workloads.jpg

When running GeekBench4 in a cloud environment such as AWS, Azure or Digital Ocean, keep in mind that Multi Core performance reported from your VM may differ from "bare metal" meaning Geekbench and the OS have access to ALL CPUs and are not sharing them with other VMs. In the cloud you may or may not be sharing CPU time with other VMs, so if you notice your multi core score does not seem to scale based on single core performance you may want to find a better fitting VM type.

Compute Tests (GPU Compute)[edit]

GeekBench4 has a brand new set of benchmarks! Woohoo! Now you can use Geekbench to test out GPU performance, specifically CUDA and OpenCL workloads. You can select multiple GPUs to test (one at a time). If you have a CPU with integrated GPU you can also select and test out the performance of your integrated graphics!

Geekbench4-main-section-gpu-compute-benchmark-compute-api-cuda-opencl.jpg

  • CUDA - Is used for Nvidia Graphics cards, to my knowledge you can't use CUDA on say, an AMD GPU, but I could be wrong here.
  • OpenCL - Is used by extremely smart people, mainly C++ programmers to create applications do crazy things.

CUDA Workloads[edit]

To run GeekBench4's GPU compute test using CUDA, you must have a NVIDIA GPU with CUDA SDK installed.

geekbench4 --compute CUDA

Below is the output from GeekBench4's CUDA GPU compute test. As you can see there are multiple workloads that are benchmarked, each with it's own score. There's also a CUDA score that is an index / combined score from each workload. Higher scores are better.

Geekbench 4.0.0 Pro : http://www.primatelabs.com/geekbench/

System Information
  Operating System        Microsoft Windows 10 Home (64-bit)
  Model                   Gigabyte Technology Co., Ltd. Z87X-UD3H
  Motherboard             Gigabyte Technology Co., Ltd. Z87X-UD3H-CF
  Processor               Intel Core i7-4790K @ 3.99 GHz
                          1 Processor, 4 Cores, 8 Threads
  Processor ID            GenuineIntel Family 6 Model 60 Stepping 3
  Processor Package       Socket 1150 LGA
  Processor Codename      Haswell
  L1 Instruction Cache    32.0 KB x 4
  L1 Data Cache           32.0 KB x 4
  L2 Cache                256 KB x 4
  L3 Cache                8.00 MB
  Memory                  16.0 GB DDR3 SDRAM 666MHz
  Northbridge             Intel Haswell 06
  Southbridge             Intel Z87 C1
  BIOS                    American Megatrends Inc. 10b
  Compiler                Visual C++ 190023918
  CUDA Device Name        GeForce GTX 1070

CUDA
  Sobel                       361327         15.9 Gpixels/sec
  Histogram Equalization      230858         7.21 Gpixels/sec
  SFFT                         28019              69.8 Gflops
  Gaussian Blur               164599         2.88 Gpixels/sec
  Face Detection               47280     13.8 Msubwindows/sec
  RAW                         836802         8.10 Gpixels/sec
  Depth of Field              521083         1.51 Gpixels/sec
  Particle Physics            222920              35240.3 FPS

Benchmark Summary
  CUDA Score                  190956



Geekbench4-gpu-compute-results-cuda.jpg

  • [Sobel] - Sobel is used to process images and do crazy stuff like detect edges / features of images. There's a lot of applications and [libraries] that utilize this.
  • [Historgram Equalization] - This adjusts the contrast of images for various reasons / use cases. If you deal with images this may be of interest to you.
  • [SFFT] - This is a research project funded by DARPA that compresses radio waves and stuff....far out man.
  • Guassian Blur - PCMASTERRACE video games use this, maybe consoles do to, but the more power, the more of dat Guassian you get :)
  • [RAW] - Something to do with images and Adobe. Not a Photoshop person so not sure how this is relevant (i'm sure there is one)
  • Depth of Field - PCMASTERRACE video games use this, maybe consoles do to, but the more power, the more of dat DoF you get :)
  • Particle Physics - PCMASTERRACE video games use this, maybe consoles do to, but the more power, the more dem particles you get :)

OpenCL Workloads[edit]

To run GeekBench4's GPU compute test using OpenCL (AMD and Nvidia and Intel) run the following

geekbench4 --compute OpenCL


The exact same workloads as CUDA are ran for OpenCL, the only difference is the API / SDK used. If you want to compare Nvidia against AMD then using the OpenCL version would be ideal as you can compare apples to apples.

Geekbench 4.0.0 Pro : http://www.primatelabs.com/geekbench/

System Information
  Operating System        Microsoft Windows 10 Home (64-bit)
  Model                   Gigabyte Technology Co., Ltd. Z87X-UD3H
  Motherboard             Gigabyte Technology Co., Ltd. Z87X-UD3H-CF
  Processor               Intel Core i7-4790K @ 3.99 GHz
                          1 Processor, 4 Cores, 8 Threads
  Processor ID            GenuineIntel Family 6 Model 60 Stepping 3
  Processor Package       Socket 1150 LGA
  Processor Codename      Haswell
  L1 Instruction Cache    32.0 KB x 4
  L1 Data Cache           32.0 KB x 4
  L2 Cache                256 KB x 4
  L3 Cache                8.00 MB
  Memory                  16.0 GB DDR3 SDRAM 666MHz
  Northbridge             Intel Haswell 06
  Southbridge             Intel Z87 C1
  BIOS                    American Megatrends Inc. 10b
  Compiler                Visual C++ 190023918
  OpenCL Platform Vendor  NVIDIA Corporation
  OpenCL Platform Name    NVIDIA CUDA
  OpenCL Device Vendor    NVIDIA Corporation
  OpenCL Device Name      GeForce GTX 1070

OpenCL
  Sobel                       352803         15.5 Gpixels/sec
  Histogram Equalization      228232         7.13 Gpixels/sec
  SFFT                         27501              68.6 Gflops
  Gaussian Blur               183022         3.21 Gpixels/sec
  Face Detection               39745     11.6 Msubwindows/sec
  RAW                         1343701         13.0 Gpixels/sec
  Depth of Field              629236         1.83 Gpixels/sec
  Particle Physics             47413               7495.4 FPS

Benchmark Summary
  OpenCL Score                168367


The workloads for OpenCL are the same as CUDA, interestingly there seems to be a difference in performance based on the Compute API you choose, if you can use either CUDA or OpenCL seems like you should pick the API that's fastest :)

Geekbench4-gpu-compute-results-OpenCL.jpg

List Available GPU devices to benchmark with GeekBench4[edit]

To view a list of available GPUs to benchmark with GB4, run the command below which uses the --compute-list option for GeekBench4

geekbench4 --compute-list

The output of the command will display the device and software available to use. In this case my GTX 1070 can be tested with the CUDA SDK as well as OpenCL.

Geekbench 4.0.0 Pro : http://www.primatelabs.com/geekbench/

CUDA
0 0 GeForce GTX 1070
OpenCL
0 0 GeForce GTX 1070

Uploading Geekbench results to online browser[edit]

You can upload both the CPU and GPU Compute tests to Geekbench's online browser that stores results from many sources. To upload these results, see image below.

Geekbench4-cpu-results-upload-geekbench-browser.jpg