While it may seem NVIDIA is in the middle of transition to Maxwell, they haven’t given up on Kepler yet. Latest, and perhaps the last, in the Kepler lineage is Tesla K80, featuring new silicon.
Tesla K80 is designed with the singular goal of offering the maximum throughput within 300W. As a result, K80 is a dual-GPU product based on a refreshed GK210 chip. It is largely similar to its predecessor, GK110, also manufactured on TSMC’s mature 28nm, with likely under-the-hood improvements made towards power and efficiency. Also, register file size and L1 cache are doubled. To get these two massive chips within the magic 300W, the K80 GPUs feature 2 x 13 SMs (2 are disabled) are declocked to only 562 MHz, though featuring a boost clock – a novelty for Tesla GPUs – upto 875 MHz for less strenuous workloads. The K80 is pared with a whopping 2 x 12 GB of 5 GHz GDDR5 RAM.
The end result is a staggering 8.74 TFLOPS and 2.91 TFLOPS of single and double precision throughput respectively, and sets the record within 300W. K80 is being aggressively targeted at supercomputing and primarily science and engineering applications.