AMD FireStream was AMD’s brand name for its Radeon product line based on GPGPU and / or stream processing orientation on supercomputers. Originally developed by ATI Technologies on the entire Radeon X1900 XTX in 2006, the product line was previously qualified as both ATI FireStream and AMD STREAM processor. The AMD FireStream can also be used as a floating point co-processor for downloading calculations from the CPU, which is part of the Torrenza initiative. The FireStream line has been discontinued since 2012, when the GPGPU workloads folded completely into the AMD FirePro line. Content
1 General information 2 History 3 Models 4 Software 4.1 Software Development Kit 4.1.1 AMD APP SDK 5 Reference points 6 Limitations 7 See also 8 References 9 External links
The FireStream line is a series of add-ons on expansion cards released from 2006 to 2010, based on standard Radeon GPUs, but designed to serve a general co-processing purpose, rather than the rendering and output of 3D graphics. Like the FireGL/FirePro line, they were given more memory and memory bandwidth, but the FireStream cards do not necessarily have video output ports. All support for 32-bit single precision floating point, and all but the first 64-bit double precision version support. The line partnered with new APIs to provide superior performance to the existing OpenGL and Direct3D shading APIs could provide, starting with the near-metal, followed by the OpenCL and Computing Stream SDK, and finally integrated into the App SDK.
For highly parallel floating point arithmetic workloads, the cards can accelerate large computations by more than 10 times; Folding @ Home, the oldest and one of the most visible users of the GPGPU, gained 20-40 times the CPU performance. Each pixel and vertex shading, or unified shading in later models, can perform arbitrary floating-point calculations. History
Following the release of the Radeon R520 and GeForce G70 GPU cores with programmable shaders, high floating point performance caught the attention of academic and commercial groups, experimenting with the use of continuation for non-graphics work. Interest led ATI (and Nvidia ) to create GPGPU products - capable of massively parallel computing general purpose mathematical formulas - to process the heavy calculations traditionally performed on CPUs and specialized floating point co-processors. GPGPUs are projected to have immediate performance improvements of a factor of 10 or more, compared to more contemporary multi-socket-only CPU computing.
With the development of the high performance XFX X1900 nearly complete, ATI based its first mainstream processor design on it, announcing it as the next ATI FireStream along with the new Fence Metal API at SIGGRAPH 2006. The core itself was almost unchanged, except for doubling onboard memory and bandwidth, similar to the FireGL V7350; new driver and software support make up most of the difference. Folding@Home began using the X1900 for general computing, using a pre-release version 6.5 of the ATI Catalyst driver, and reported 20-40x improvement in the GPU over the CPU. The first product was released in late 2006, renamed the AMD STREAM processor after the AMD merger.
The brand became AMD FireStream with the second generation flow processor in 2007, based on the RV650 chip with new unified shaders and double precision support. Asynchronous DMA has also improved performance by allowing a larger memory bank without the aid of the CPU. One model was released, the 9170, for the initial price of $99. Plans include the development of an MXM module flow processor by 2008, for mobile computing, but it was never released.
The third generation quickly followed in 2008 with performance improvements from the RV770 core; the 9250 had nearly twice the performance of the 9170, and became the first single chip teraflop processor, despite dropping the price to less than $1000. A faster sibling, the 9270, was released shortly thereafter, for $99.
In 2010, the final generation of FireStreams came out, the 9350 and 9370, cards based on the Cypress chip that appeared in the HD 5800. This generation doubled the performance of the previous generation, to 2 teraflops in the 9350 and 2.6 teraflops in the 9370, and was the first built from scratch for OpenCL. This generation was also the only one equipped with a completely passive coolant, and active cooling was not available.