The world of server-class graphics cards just got a lot more interesting. Today, AMD introduced the FirePro S10000, an astoundingly powerful graphics card with an unmatched credential: It is the first commercially available server-class card with over one teraflops (floating-point operations per second) of double-precision floating-point performance. On top of that, it has almost six teraflops of single-precision floating-point performance. This is a monster of a card that will undoubtedly turn some heads in the high-end market — for performance reasons, of course, and because a single S10000 will set you back $3600.
Impressively, the Tahiti-based FirePro S10000 isn’t just about raw horsepower. Its performance per watt is 3.94 gigaflops. AMD claims that’s up to 4.7 times better than the dual-GPU Nvidia Tesla K10. Things are going to heat up yet further when Nvidia releases itsTesla K20 this month. Nvidia’s upcoming GPU release is also interesting because it breaks the teraflops barrier on double-precision floating-points. AMD’s offerings are superior on paper: 1.48 double-precision and 5.91 single-precision compared to 1.17 double-precision and 3.52 single-precision for Nvidia’s GPU. Until they’re both available for benchmarking in real-world scenarios, it isn’t clear if the numbers will make much of a difference.
The S10000 packs an impressive 6GB of GDDR5 RAM with 480GB/s of bandwidth shared between the two GPUs. It sports four mini DisplayPorts and a single DVI port. Its max power draw is 375 watts, but it will need a power supply of at least 750 watts. It uses PCI Express x16 3.0, and takes up two slots. Since this is a server-class card, AMD makes absolutely no bones about it being a full height/full length card. Official driver support for Windows Server 2008 R2 SP1, Windows 8, Windows 7, Windows Vista, and both 32 and 64-bit Linux are available. Windows Server 2012 support is planned for a later release. As one would expect, it comes with support for DirectX 11, OpenGL 4.2, and OpenCL 1.2.
While it seems like an inconspicuous check mark on the spec sheet, OpenCL is really the tech that gives software developers the ability to easily take advantage of the insane parallel-computing technology locked away in these server- and workstation-class graphics cards. Instead of bogging down the CPU with these tasks, OpenCL is a layer that allows software developers to tap into the GPU’s potential without being an expert at exploiting the complexities of the system. While GPUs aren’t perfect for every situation, they are incomparably powerful when it comes to certain high-end workloads.
Most of the time, we think about graphics processors drawing windows and making complex 3D worlds. That’s not all they can do, and that’s why these big increases in power are a big deal. Researchers can use their GPUs to finish processing data in a fraction of the time it would take with CPUs. Massive server farms at companies like Google can reduce their power and si e footprint simply by having a single server do the task of many servers. The folks doing extreme renderings for CGI movies can create and use custom software that uses the GPU’s horsepower without having to start from scratch. OpenCL technology in combination with ridiculously powerful graphics cards like the S10000 and Tesla K20 are opening all sorts of new doors, and that is something to get excited about.
AMD launches Tahiti-based FirePro S10000, the first teraflops graphics card
0 comments:
Post a Comment