Geekbench: The way it truly works


Benchmarking smartphones and computer systems includes testing machine efficiency in a couple of totally different classes. You’ll be able to evaluate numerous totally different metrics, together with graphical efficiency, synthetic intelligence, and computation. Geekbench is a benchmark that has change into a staple of the benchmarking world, and it is primarily centered on computation. Geekbench 6 is the newest model, however what precisely is it? What does it take a look at, and the way?

What’s Geekbench?

Geekbench is a cross-platform benchmarking software that may attribute a rating to each the single-core and multi-core computational capabilities of your machine. This rating can be utilized as some extent of comparability towards adjoining units and is calibrated towards a baseline rating of two,500, which Primate Labs says is the rating of a Dell Precision 3460 with an Intel Core i7-12700. Admittedly, searching by way of Geekbench 6 scores appears to level to that individual CPU solely barely scraping 2000 factors in single-core, however apart from that, the premise is {that a} machine that scores 5000 is claimed to have twice the efficiency of the i7-12700.


Within the case of Geekbench 6, it is the newest iteration of the Geekbench benchmarking suite, and it goals to measure your smartphone’s capabilities within the ways in which truly matter in the case of utilizing any of the finest telephones.

  • Larger photographs in resolutions captured by fashionable smartphones (12-48MP)
  • HTML examples consultant of recent net design requirements
  • A bigger library of photographs for import assessments
  • Bigger maps for navigation assessments
  • Larger and extra fashionable PDF examples
  • A rise in Clang workload measurement

There are GPU computation assessments too, and it could possibly take a look at OpenCL, Steel, and Vulkan. The GPU computation benchmark makes use of machine studying workloads equivalent to background blur and face detection to check object recognition capabilities. On high of that, it runs picture modifying workloads, equivalent to horizon detection, edge detection, and Gaussian blur. Lastly, there are picture synthesis workloads that execute function matching and stereo matching, together with a simulation benchmark that simulates particle physics.

What platforms does Geekbench 6 assist?

Geekbench 6 helps the next platforms, with Home windows on Arm assist for a few of the finest laptops set to come back with Geekbench 6.1:


Minimal model




Android 10

AArch64, x64


iOS 15



Ubuntu 18.04 LTS

AArch64, x64

CentOS, RHEL, extra to come back


macOS 11

AArch64, x64

Home windows

Home windows 10


AArch64 coming with Geekbench 6.1

How does Geekbench’s CPU benchmark work?

The Snapdragon 8 Gen 2 (proper) in comparison with the Snapdragon 8 Gen 1 on Geekbench. 

Geekbench’s CPU benchmark is damaged down into a lot of key assessments which have a single-core and multi-core part. Every part is grouped into two subsections: integer workloads and floating level workloads. There are gaps between every take a look at of two seconds by default to reduce the impact that thermal points have on efficiency.

Geekbench 6 launched a shared activity mannequin for multithreading testing, the place workloads are shared throughout a number of threads so as to characterize extra real looking workloads. Beforehand, Geekbench unfold workloads throughout particular person threads, which scales nicely however provides little or no inter-thread communication. Within the case of shared activity fashions, every thread processes half of a bigger shared activity. It does not scale as nicely however is extra consultant of real-world use circumstances.

Scores are computed utilizing a weighted arithmetic imply of subsection scores, with the integer subsection accounting for 65% of the rating and the floating level subsection accounting for the remaining 35%.

As for a way Geekbench assessments the flexibility of your machine’s chipset, it assessments several types of workloads break up into classes. These classes are break up into productiveness, developer, machine studying, and picture synthesis.

Geekbench 6 Productiveness workloads

These are workloads that take a look at how performative your machine is at on a regular basis vital duties.

File compression

File compression workloads take a look at how good your machine is at compressing and decompressing recordsdata utilizing totally different compression codecs. It fashions use circumstances the place a person could look to compress a file to ship to another person so as to cut back knowledge and bandwidth. It compresses the Ruby 3.1.2 supply archive, which is a 75MB archive containing 9,841 recordsdata, utilizing LZ4 and ZSTD compression. It then verifies the compressed recordsdata through an SHA-1 hash.

These recordsdata are then saved utilizing an in-memory encrypted file system, and this workload makes use of directions that speed up AES encryption and decryption. It additionally makes use of directions that speed up SHA-1 hashing algorithms.


We use navigation throughout all kinds of units, notably smartphones. The navigation workload goals to generate instructions between a sequence of places, and fashions folks utilizing apps like Google Maps in offline mode. It makes use of Dijkstra’s algorithm to calculate 24 totally different routes on two totally different OpenStreetMap maps. One is in Waterloo, Ontario, and one is in Toronto, Ontario.

HTML5 Browser

The HTML5 browser opens a lot of HTML5 pages and fashions a person searching the online in a contemporary browser equivalent to Chrome or Safari. It makes use of a headless browser and opens, parses, lays out, and renders texts and pictures based mostly on in style websites, together with Instagram, Wikipedia, and Ars Technica. It makes use of the next libraries:

  • Google Gumbo because the HTML parser
  • litehtml because the CSS parser, format, and rendering engine
  • FreeType because the font engine
  • Anti-Grain Geometry because the 2D graphics rendering library
  • libjpeg-turbo and libpng because the picture codecs

This take a look at renders eight pages in single-core mode and 32 pages in multi-core mode.

PDF Render

The PDF render workload opens complicated PDF paperwork utilizing PDFium, which is Chrome’s PDF renderer. It renders PDFs of park maps from the American Nationwide Park Service, with sizes starting from 897kb to 1.5MB. These recordsdata include giant vector photographs, traces, and textual content.

This take a look at renders 4 PDFs in single-core mode and 16 PDFs in multi-core mode.

Photograph Library

The photograph group workload categorizes and tags photographs based mostly on objects they include, permitting customers to look their photographs by key phrase in picture organizer apps. It makes use of MobileNet 1.0 to categorise photographs and a SQLite database to retailer the photograph metadata alongside their tags.

This workload performs the next steps for every photograph:

  1. Decompress the photograph from a compressed JPEG file.
  2. Retailer photograph metadata in an SQLite database. This database is pre-populated with metadata for greater than 70,000 photographs.
  3. Generate a preview thumbnail and encode it as a JPEG.
  4. Generate an inference thumbnail.
  5. Run a picture classification mannequin on the inference thumbnail.
  6. Retailer picture classification tags in an SQLite database.

The photograph library workload operates on 16 photographs in single-core mode and 64 photographs in multi-core mode.

Geekbench 6 Developer workloads

Developer workloads in Geekbench 6 measure how nicely your machine handles typical developer duties equivalent to textual content modifying, code compilation, and asset compression.


The Clang compiler is used to compile the Lua interpreter, modeling the use case of builders constructing their code and just-in-time compilation that customers ceaselessly will expertise on their units. It makes use of musl libc because the C customary library for the compiled recordsdata. It compiles eight recordsdata in single-core and 96 recordsdata in multi-core mode.

Textual content processing

Textual content processing hundreds quite a few recordsdata, parses their contents utilizing common expressions, shops metadata in an SQLite database, and exports the content material to a unique format. It fashions typical textual content processing algorithms that manipulate, analyze, and remodel knowledge for publication and perception achieve.

This workload is applied in a mix of Python and C++, utilizing Python 3.9.0 and processing 190 markdown recordsdata for enter.

Asset compression

Asset compression compresses 3D textural and geometric property utilizing quite a lot of in style compression codecs equivalent to ASTC, BC7, and DXT5. It fashions the usual content material compression pipelines which might be utilized by recreation builders.

The workload makes use of bc7enc for its BC& and DXTC implementations, and Arm ASTC Encoder for its ASTC implementation.

Machine Studying workloads

Machine studying workloads primarily measure how nicely your CPU can deal with recognizing objects in photographs and scenes.

Object detection


The article detection workload makes use of machine studying in order that it could possibly detect and classify objects in photographs. It makes use of a convolutional neural community referred to as MobileNet v1 SSD to detect and classify objects in photographs, and the photographs are sized at 300×300 pixels. It performs the next steps to determine objects in a picture:

  1. Load the photograph
  2. Extract objects from the photograph utilizing MobileNet v1 SSD
  3. Generate a confidence or detection rating that represents the accuracy of the detection
  4. Draw a bounding field across the object and output a confidence rating

Object detection processes 16 photographs in single-core mode and 64 photographs in multi-core mode.

Background blur

The background blur workload separates the foreground from the background in video streams and blurs the background, identical to what companies equivalent to Zoom, Discord, and Google Meet can do.

Picture Enhancing

Picture modifying workloads measure how nicely your CPU can deal with each easy and complicated picture edits.

Object remover

The article remover workload removes objects from photographs and fills within the hole left behind, modeling a content-aware fill and Google’s personal Magic Eraser. The workload is offered a 3MP picture with an undesirable area, and the workload removes this area and makes use of an inpainting scheme to reconstruct the hole that’s left behind.

Horizon detection

The horizon detection workload can detect and straighten uneven or crooked horizon traces so as to enhance photographs. It fashions horizon line correctors in photograph modifying apps and makes use of the Canny edge detector to use a Hough remodel to detect the horizon line. It makes use of a 48MP photograph as its enter.

Photograph filter

The photograph filter workload applies filters to reinforce photograph look, modeling frequent filters in social media apps like Instagram. It applies the next results to 10 totally different photographs, with photographs ranging in measurement from 3MP to 15MP.

  • Shade and blur filters
  • Stage changes
  • Cropping and scaling
  • Picture compositing


The HDR workload blends six common photographs to create a single HDR photograph that’s colourful and vibrant. It fashions HDR options present in fashionable smartphone digicam apps, making a single 16MP HDR picture from six 16MP common photographs.

Picture Synthesis

These workloads measure how your CPU can deal with creating totally synthetic photographs.

Ray tracer

Ray tracing is all the fashion, and it may be used to generate photorealistic photographs by modeling how mild rays work together with objects in digital scenes. This fashions the rendering processes that 3D rendering software program like Blender or Cinema 4D would use.

Construction from movement

Construction from movement is a method that generates 3D geometry from a number of 2D photographs. Augmented actuality programs use methods like these to grasp real-world scenes. This workload takes 9 2D photographs of the identical scene and constructs an estimate of the 3D coordinates of the factors which might be seen in each photographs.

The way to obtain Geekbench


Geekbench is likely one of the go-to benchmarks that individuals use to check units like the very best telephones, laptops, and tablets, and you may obtain it from the Apple App Retailer, the Google Play Retailer, and the Primate Labs web site.

Supply hyperlink


Please enter your comment!
Please enter your name here