Even after the introduction of the Nvidia GM204 with the Maxwell architecture, the huge circuit GK110 continues to be one of the most complex of all time, with a size of 561 mm2 and 7.1 billion transistors. Despite a new architecture, Kepler is not ready to retire, but is reborn with GK210.
The newcomer is similar to the GK110 with 2,880 CUDA cores, but each SMX cluster with 192 CUDA cores has a doubled cache of 512 KB (register) and 128 KB (shared memory). These changes will enhance the performance of general calculations and also make the circuit more energy efficient than its predecessor, even though it is manufactured on the same 28-nanometer technology.
The newly launched Tesla K80 calculation card comes with two GK210s, each of which has two SMX clusters suspended. It provides 2 x 2496 CUDA cores clocked at 562 MHz at base and 875 MHz at boost. This gives a computing power of up to 8.74 TFLOPS at single precision and 2.91 TFLOPS at double precision.
This includes two 384-bit memory buses with a full 12 GB of GDDR5 memory each, ie a total of 24 GB on a graphics card. The memory comes clocked at 5,000 MHz for a bandwidth of 240 GB / s per graphics circuit. All this can be accommodated thanks to the graphics card’s dynamic clock frequency in a TDP value of 300 W and comes with passive cooling, intended to be cooled by the air flow from external fans.
The Nvidia Tesla K80 will be available at retailers and partner manufacturers shortly for around $ 5,000, corresponding to approximately SEK 46,300 including VAT.