The industry-leading Cypress GPUs (in the Radeon 5870 and new FirePro line) have made their way into AMD’s FireStream GPU co-processors. These “compute accelerators” are intended for highly parallel, compute-intensive workloads in scientific, financial and academic arenas. Unlike the discrete FirePro/Radeon graphics cards, the FireStream compute accelerators use a passive heat sink so they can slide into rack mounted HPC servers and expansion systems for x64 systems.
From the press release: The AMD FireStream 9350 delivers 2.0 TFLOPS of single precision performance and 400 GFLOPS of double precision floating point performance in a single-slot, 150W solution with 2GB of GDDR5 memory, enabling breakthrough compute density (almost double the previous generation). The AMD FireStream 9370 delivers up to 2.64 TFLOPS of single precision performance and 528 GFLOPS of double-precision performance, and includes 4GB of high-speed GDDR5 memory, at a maximum board power of 225 watts. In addition, the AMD FireStream 9350 and 9370 both support leading industry standard application interfaces, including OpenCL, DirectX 11 and OpenGL.
The FireStream boards are AMD’s higher performance answer to Nvidia’s Tesla Fermi-based co-processors. The higher floating-point performance aside, with the FireStream, AMD is also focused on delivering a solution based on open standards like OpenCL rather than the proprietary solutions like Cuda.
AMD announced the ATI Stream v2.1 SDK (AMD's OpenCL computing implementation to accelerate application performance using both the CPU and GPU resources in a systems).
New features include:
Support for OpenCL / OpenGL interoperability, to reduce the overhead of passing data for display purposes, enabling a richer and more responsive visual experience for the user.
Support for OpenCL images, providing developers with access to hardware-accelerated texture features on AMD GPUs.
OpenCL extension support for AMD media operations in OpenCL, giving developers a set of OpenCL kernel operations commonly used in multimedia applications.
Support for OpenCL byte addressable stores allowing more natural and efficient code for applications, such as image processing, that depend on the ability to update data at smaller than 32-bit granularities.
OpenCL extension support for device fission in OpenCL, enabling developers to sub-divide an OpenCL device and allowing multiple work kernels to be assigned to that device.
Integration of Stream KernelAnalyzer 1.5 installer, which helps developers to statically analyze OpenCL kernel performance on AMD graphics processors.
Support for next-generation ATI FirePro professional graphics card family, including the ATI FirePro V8800, and the latest ATI Radeon and ATI Mobility Radeon graphics cards from AMD (List of supported hardware).
AMD’s newly released Maya Tessellator Plug-in allows users of Autodesk’s Maya to take advantage of FirePro Graphics GPU tessellation hardware. The readme.doc file in the package describes how to install and use the plug-in. Binaries are included for Windows XP 32 and Windows XP 64 operating systems Unfortunately you must create a free AMD developer account to download the plug-in.
AMD and announced a joint development agreement as part of the AMD effort to greatly expand the use of real-time physics with graphics through the open source Bullet Physics engine using OpenCL and/or DirectCompute in DirectX 11.
A great quote from the release: “Proprietary physics solutions divide consumers and ISVs, while stifling true innovation; our competitors even develop code that they themselves admit will not work on hardware other than theirs”
I’ve been waiting for this video from Siggraph 2009! Rick Bergman (Senior VP and General Manager of AMD), discusses how computer graphics technology can enhance Hollywood productions. The video is naturally about the power of the GPU and in particular how it enables real time creativity. There are some great clips of MachStudio Pro including hardware tesselation for displacement mapping.
I can’t believe I was out of town for AMD’s big Sept 10 event. They announce their Eyefinity multi-display technology (driving up to six monitors simultaneously at resolutions up to 2,560x1,600 pixels each), and they demod Crysis, the standard gaming benchmark for high-end 3D hardware, running on an iPhone. Huh?
At the event, AMD unveiled its next-generation GPU architecture with 2.5 teraFLOPS of floating-point power (over twice current high-end cards!). To show off how this could translate to applications, they showed OTOY’s software, running on AMD servers and new GPUs, delivering 3D games in real-time over the Internet to the iPhone. The report from Ars Technica is impressive and worth the read.
MacResearch has posted two video tutorials on OpenCL - what is it and how to code with it. While the examples are for given for the Mac OS, OpenCL is a cross platform standard and will run on any modern CPU or GPU (or both) with drivers. AMD and Nvidia will have OpenCL drivers for their GPUs under Windows and Linux. AMD and Intel will support OpenCL on their CPUs. AMD has already shipped its first OpenCL implementation for its Athlon and Opteron processors as a free download as part of the ATI Stream SDK v2.0 Beta Program.
I’ve seen the Froblins demo before with 3000 creatures scurrying around the screen. But always thought this was primarily a hardware tesselation and rendering demos. Turns out it is also about concurrently running interactive Artificial Intelligence code on entirely on the GPU. The full story is parallel artificial intelligence computation for dynamic pathfinding and local avoidance on the GPU, massive crowd rendering with LOD management with high-end rendering capabilities such as GPU tessellation for high-quality close-ups and stable performance, terrain system, cascaded shadows for large-range environments, and an advanced global illumination system.
Yesterday I posted a video showing OpenCL scaling across 24 CPU cores. The same demo was actually being shown at SIGGRAPH scaling across 12 CPU cores. We got a capture of the demo, but the most important thing about it is the section where the presenter explains about why OpenCL is is so significant and why AMD is showing it running on the CPU and not just talking about running it on the GPU (like everyone else). OpenCL lets you write data-parallel algorithms data-parallel processing (for scientific computations, video and image processing, game AI and physics, etc.) that spread themselves automatically across available resources, be they CPU or GPU.
See the video below, but here is the transcript:
“The important thing to note about OpenCL, is that it is not simply about running on the GPU. OpenCL is about running on heterogeneous systems - ALL the processors in your system. With AMD’s OpenCL implementation, you will be able to take one source code base and re-target it to your CPUs or GPUs - it will run on both - and take advantage of your entire platform. ”.
My comment: Now that (in addition to the whole open standards thing) is why I am such a big champion of OpenCL.
AMD is offering a free OpenCL for CPU beta download as part of the ATI Stream SDK v2.0 Beta Program. The Stream SDK allows you to take advantage of GPGPU computing and develop your applications in a high-level language - OpenCL. It allows you to divide software workloads between different hardware elements such as multi-core CPUs, GPUs, and DSPs.
Below is a video of a 4P six-core AMD Opteron processor-based system (24 total cores) running an OpenCL-based, fluid/particle simulation. The demo runs through enabling progressively more CPU cores and you see the performance scaling. Impressive.
Normally you think of just the GPU for this kind of computational processing. But reality is you need both the GPU and CPU for optimal performance and OpenCL provides this. I stole the quote form an AMD blog:
“Of course no application runs entirely on the GPU. Beyond the obvious need for CPUs to drive execution, most mainstream applications are heterogeneous in nature. They have some functions that accelerate well on multicore CPUs, and others that are perfectly suited for a GPU’s data parallel architecture. A good development platform needs to take that into account - this is the difference between GPGPU as a niche accelerator and GPGPU as a new baseline feature, ready for tomorrow’s systems and applications.”
FireUser.com is a community resource for visualization, 3D, video and engineering professionals to learn about the latest acceleration and display technologies, discuss support issues, as well as influence the features and direction of the FireGL and FirePro accelerator line.
AMD and 
I can’t believe I was out of town for AMD’s big Sept 10 event. They announce their 
