Heterogeneous System Architecture

from Wikipedia, the free encyclopedia

The Heterogenous System Architecture ( HSA for short , formerly Fusion System Architecture ) is a processor concept developed by AMD to combine the main and graphics processor as efficiently as possible on one computer chip ( die ) and thereby accelerate the execution of specialized programs. Among other things, this should significantly increase the overall performance of APUs ( Accelerated Processing Unit ). The main feature of HSA is that the CPU cores and the graphics processor access a shared address space and memory (RAM).


AMD announced the HSA program at the AFDS ( AMD Fusion Developer Summit ) 2011, with the aim of immensely accelerating the processing speed of various tasks. Due to trademark problems with Arctic (known among other things for CPU and GPU coolers), AMD had to rename the program called FSA ( Fusion System Architecture ) to HSA.

At the AFDS 2012 , AMD announced the establishment of the so-called HSA Foundation , where AMD wanted to work with other chip developers on a common HSA concept. Members of the HSA Foundation include ARM and Texas Instruments .

In the future, AMD plans, in cooperation with ARM, to encourage the software industry to optimize their products for HSA and possibly also to gain several hardware manufacturers.

Founder of the HSA Foundation


AMD graphics stack under Linux with HSA (
using the amdkfd driver )

The concept of the HSA is based on the fact that the graphics part of APUs can take over general tasks (GPGPU) and thereby accelerate special computing operations. The graphics unit can be programmed with C ++ and OpenCL . The "HSA SDK" development package should also offer the "HSA intermediate language" (HSAIL) as an abstraction level in order to optimize C ++ code for the APU. This intermediate layer also maintains the compatibility between different APUs.

The currently highest expansion level of HSA in the desktop and notebook market is the Kaveri APU , which can guarantee coherence in the memory and which also masters AMD's "Heterogeneous Uniform Memory Access" (HUMA).

The GPU is used because, with its parallel execution units, it is very good at performing floating point arithmetic operations quickly and efficiently. In contrast to the normal CPU, these offer a lot more floating point units. Current CPUs only offer one floating point unit per module at AMD ( AMD Bulldozer architecture) and at Intel only one floating point unit per core (Intel Sandybridge architecture).


The characteristics of a processor with HSA support are that the graphics processor and main processor are combined on the same computer chip. This facilitates the faster data exchange between the two units, and thus tasks can be accelerated considerably on the software side. Among other things, this reduces the complexity of computer motherboards, as no extra graphics chip needs to be soldered on there. Another important feature is that in the final expansion stage the graphics part has a coherent memory and shared address space and can be programmed with C ++ or OpenCL. AMD has already integrated this into its Kaveri APUs with Graphics-Core-Next graphics architecture. Chip manufacturers can also integrate 3rd party IP components into their microchips via the HSA . This is the case with the successor generation of the energy-saving ("low power") chips based on the Jaguar architecture. For example, the ARM TrustedZone is integrated directly into the x86 chip.

Further development

AMD plans to use various logic units from third parties such as Ex. ARM cores add their APUs and these therefore specifically adapted to the purpose. The third expansion stage of HSA is represented by the Kaveri and Godavari models, which offer coherent memory and a common address space for CPU and GPU. They are equipped with 2–4 Steamroller cores (successor to Piledriver ) and Graphics Core Next graphics part (currently the latest generation of graphics chips from AMD).

Sony announced in February 2013 that it would use x86 APUs with AMD's Jaguar as CPU and "Graphics Core Next" as GPU architecture in their PlayStation 4 . Like Kaveri and Godavari, these have the third HSA expansion stage and, as Kabini and Temash , have replaced the previous models based on the Bobcat architecture in the low-end desktop and notebook market and in the tablet market.

The last HSA expansion stage for the time being are the Carrizo APUs with Excavator cores, which are to be supplemented by Bristol Ridge versions for the desktop in 2016 .


In terms of software, HSA is already supported by some large software houses. The supported programs include the well-known image processing programs Gimp , Adobe Photoshop and Darktable , where various functions run on the graphics part via OpenCL. Furthermore the office package LibreOffice (in the spreadsheet), various software for converting and displaying videos such as B. Handbrake , x264 , Sony Vegas , Adobe Premiere , Final Cut Pro , Avisynth as well as web browsers and software for 3D rendering (e.g. Blender , Maya , 3dsMax ).

Software programmed with OpenCL is accelerated on graphics processors from all manufacturers. In comparison, however, NVIDIA's in-house development CUDA falls behind .

PGI (The Portland Group) plans to develop a corresponding HSA compiler in partnership with AMD, with which programs can be compiled without much effort on your own, which are accelerated by the HSA extensions.

Processors with HSA support

Overview of processors with HSA support
developer CPU architecture GPU architecture Instruction set Code name Release date production
AMD AMD Jaguar Graphics Core Next x86 ( AMD64 ) PlayStation 4 SoC 2013 28 nm
AMD AMD Steamroller Kaveri 2014
AMD AMD Puma + Mullins & Beema
AMD ARM Cortex A57 ARMv8 (64-bit) Seattle
AMD AMD Steamroller x86 ( AMD64 ) Godavari 2015
AMD AMD Excavator Carrizo
AMD AMD Puma + Carrizo-L
AMD AMD Excavator Bristol Ridge 2016 28 nm
AMD AMD Zen Summit Ridge 2016 14 nm
AMD AMD K12 ARMv8 (64-bit) n / A
AMD AMD Zen 2 x86 ( AMD64 ) / 2019 7 nm

Web links

Individual evidence

  1. AMD's "Fusion" -based architecture to be renamed - Article on ComputerBase , January 20, 2012. Retrieved July 26, 2012.
  2. Martin Fischer: AFDS 2012: AMD and ARM go hand in hand. In: heise online . June 13, 2012.
  3. Martin Fischer: AFDS 2012: AMD announces APU combination processors with ARM cores. In: heise online . June 14, 2012.
  4. ^ AMD 2012–2013 Mobile Roadmap. In: CPU-World.com (English).
  5. http://www.kitguru.net/components/cpu/anton-shilov/amd-preps-bristol-ridge-apus-carrizo-for-desktops/
  6. Frequently asked questions about Photoshop and GPUs. Retrieved July 31, 2016 .
  7. LibreOffice Wiki: OpenCL Settings. May 7, 2015, accessed July 31, 2016 .
  8. GPU acceleration. (No longer available online.) Archived from the original on August 8, 2016 ; accessed on July 31, 2016 (English). Info: The archive link was inserted automatically and has not yet been checked. Please check the original and archive link according to the instructions and then remove this notice. @1@ 2Template: Webachiv / IABot / trac.handbrake.fr
  9. AMD Radeon ProRender. Accessed July 31, 2016 .
  10. http://www.planet3dnow.de/cgi-bin/newspub/viewnews.cgi?category=1&id=1351715019
  11. http://www.planet3dnow.de/vbulletin/showthread.php?t=408921