At the recent International Conference for High Performance Computing, NVIDIA unveiled the successor to its flagship Tesla K40, the K80; a dual-GPU graphics monstrosity designed to accelerate high performance computing (HPC) tasks.
The Tesla K40 was a beast in its own right, and was the first Tesla offering from the company to ship with all 2880 CUDA cores of the GK110B ‘Kepler’ architecture enabled. This time around, the K80 is based on a new GPU (GK210) with 2496 CUDA cores clocked at 562MHz that can boost up to 875MHz on the fly. The card also has 12GB of VRAM with a 384-bit memory bus width. With a total transistor count of 7.1 billion, the card draws about 300W of power. Remember that this is a dual-GPU chip, so that’s a total of 4992 CUDA cores and 24GB of VRAM.
All that power has been crammed into a double-slot card that is passively cooled. The K80 also supports Quad-SLI so four of them could be hooked up together for even crazier performance. Doing so would result in about 17.5 TFLOPS of single precision computing power, so for those that need serious computing power but don’t have the space to host and cool a large number of servers, this could be a viable alternative option.
For a card that delivers a whopping 74% improvement over its predecessor, one would expect it to at least double the $5499 price of the K40, but that’s not the case. NVIDIA is pricing the K80 at $5000, that’s a cheaper price, for a lot more performance. Although the reason behind the price drop could be the cards 300W power consumption rate, compared to the K40 that requires only 235W.
The K80 is now available for purchase and will be added as options when customizing servers from OEM’s including Cray, Dell, HP, and Quanto Computer among others.