NVIDIA's top offering for servers and workstations are the Kepler GK110-based K20 and K20X. The K20X is capable of peaking at 1.31 Teraflops for double-precision floating point math or 3.95 Teraflops when doing single-precision floating point operations. The K20 meanwhile can achieve 1.17 and 3.52 Teraflops for double and single precision floating point, respectively. The memory bandwidth with ECC disabled for the Tesla K20X tops out at 250GB/s while packing 6GB of GDDR5 video memory. The K20X has 2688 CUDA cores on its GK110 die while the K20 has 2496 cores.
The GK110-based K20s are very impressive with more information on them being available at the NVIDIA Tesla HPC page. The NVIDIA K20 has been talked about before but now it's being re-introduced for SuperComputing 12 with the GPUs powering Titan, the world's fastest super computer out of the Oak Ridge labs. There's 18,688 NVIDIA Tesla K20X GPU accelerations and a matching number of AMD Opteron processors.
Meanwhile AMD's new wares for the super computing conference is the AMD FirePro S10000. AMD claims that this new product is the industry's most powerful server graphics card.
The FirePro S10000 advertises 5.91 Teraflops for single-precision peak performance and 1.48 Teraflops on double-precision floating-point calculations. The FirePro S10000 is based upon their "GCN" (Southern Islands) architecture and packs 6GB of GDDR5 video memory for its dual GPUs with ECC memory support. Additional information on this new AMD FirePro graphics card is available from its product page.
Both the NVIDIA Tesla K20/K20X and AMD FirePro S10000 are supported by the respective company's latest Linux binary graphics drivers. Nouveau developers have even managed to work on early NVIDIA GK110 support for their reverse-engineered open-source Linux graphics driver.