After earlier launching the first Kepler GPU chip, the GK104, Nvidia has revealed that chip’s bigger brother, the GK110. Designed primarily for supercomputers and data centers, this 7.1-billion-transistor megachip is scheduled to ship in 4Q12 in Tesla K20 add-in card. For applications that don’t require double-precision math, Nvidia also announced the Tesla K10 card, which uses two GK104 GPUs.

Like the GK104, the GK110 is fabricated in TSMC’s 28nm process. The first Kepler GPUs were introduced earlier this year, primarily for PC graphics, but Nvidia had plans to bring Kepler to high-performance computing (HPC) applications as well.

The GK110 has a total of 2,880 shader units, nearly twice as many as in the GK104. But the GK110 is more than just a supersized GK104: Nvidia has included optimizations for servers and scientific workloads. These additions include dedicated floating-point blocks for double-precision math, improved thread management, and a new virtualization capability.

The company rates the new GK110 at over one double-precision teraflops, but it has withheld the target clock frequency. The full Kepler implementation includes 15 streaming multiprocessor (SMX) blocks and six memory controllers.

At the other end of the spectrum, Nvidia also released the smallest Kepler-family GPU, the GK107. This version measures only 118mm2 and is branded the GeForce GT 640. It targets low-cost video cards (around $100) and premium notebook designs, such as the recently refreshed Apple MacBook Pro. The GK107 still provides better gaming performance than the integrated GPU in Intel’s Ivy Bridge, despite the improvements Intel has made in its GPUs. But Ivy Bridge and AMD’s Trinity processors have largely eliminated the market for GPU chips smaller than the GK107.

Nvidia will face competition in delivering teraflops-level coprocessors from both AMD’s FireStream GPU-compute accelerators and Intel’s recently announced Xeon Phi chip (based on its Knight’s Corner manycore processor). All three vendors will compete for HPC coprocessor designs and are making similar performance claims. System-design decisions will likely be based on power and cost, data for which has yet to be made public. Nvidia has made a big bet on HPC by building a dedicated chip just for this nascent low-volume market.