During the SC'14 event, where the company usually unveils its Tesla compute graphics cards, Nvidia has pushed the compute performance a bit further by unveiling the new dual-GPU Tesla K80 graphics card which is based on two GK210 Kepler GPUs and packs no less than 24GB of GDDR5 memory.


While it was generally expected that the Tesla K40, based on a fully enabled GK110 GPU would be the pinnacle of the Kepler GPU, Nvidia has decided that there is still some life left in that same GPU and has unveiled the new Tesla K80, which is based on a pair of GK210 Kepler GPUs, which are practically a variant of the GK110 GPU. From a performance point of view, the new Nvidia Tesla K80 dual-GPU graphics card is pushing the performance to a whole new level, with maximum double precision compute performance (FP64) of 2.9 TFLOPS and single precision compute performance of 8.7 TFLOPS, which makes it around 74 percent faster than the GK110 equipped Tesla K40.




The new Kepler GK210 GPU, although still manufactured on the same 28nm TSMC manufacturing process, brings a few improvements. While it only has 13 out of 15 SMX enabled on each GPU, providing a total of 4,992 (2x2,496) CUDA cores on the Tesla K80. Each GPU is clocked at 562MHz base and 875MHz GPU Boost clocks and the Tesla K80 comes with 12GB of 5.0GHz clocked GDDR5 memory per GPU, for a total of 24GB of memory, connected to a dual 384-bit memory interface, which adds up to 480GB/s of memory bandwidth (240GB/s per GPU).




As noted, the GK210 is practically the 3rd revision of the GK110 GPU and has a few improvements. The GK210 comes with a different memory subsystem for each SMX, as it got a larger 512KB register file and 128KB of shared memory, which is double the amount seen on the GK110B.

While the Tesla K40 and K20X were limited at 235W, pairing up two Kepler GK210 also pushed the power envelope to a new level as the new dual-GPU Tesla K80 is a 300W graphics card. Nvidia also managed to pack those two GPUs on a standard dual-slot Tesla graphics card form-factor and even done it in a recognizable passive cooling way.




Of course, the Tesla K80 will also feature the GPU Boost, and not just any, but the new dynamic GPU Boost which allows the GPUs to go as high as the TDP allows.

Nvidia has pretty much hard-launched the new Tesla K80, and it will be available from standard partners as individual graphics card as well as standard OEMs like HP, Quanta, Cray and Dell. While Nvidia usually does not reveal MSRP of the Tesla family, the price appears to be set at US $5000, which is actually cheaper than the launch price of the Tesla K40, US $5500.





Source: Nvidia.com.