NVIDIA has outed its most powerful GPU-based processing accelerator to-date, the Tesla K40, and it’s headed to supercomputers in 2014. Mustering up to 40x the performance of the last-gen K20X, the K40 delivers 1.43 teraflops of processing power from its 2,880 CUDA cores and 12GB of GDDR5 memory. There’ll be plenty of opportunity for it to show what it can do, too; NVIDIA has inked a collaboration with IBM that will see the two company’s silicon sharing supercomputer space from next year.
The deal will see NVIDIA supply GPU-based accelerators, which IBM will pair with its own POWER processors. Previously, those CPUs had been expected to do all the work, but IBM will also update its supercomputer software so that it can take advantage of the parallel processing CUDA is capable of.
Exactly what those purposes might be is generally something of a mystery; often, it seems like the prime intention of a new supercomputer is to (often very briefly) sit at the top of the speed rankings. In fact, NVIDIA explained to us, many moments during the day – even just when we’re checking the weather – we’re tapping into the big data processing capabilities of supercomputers.
For instance, many supercomputers are used to predict weather patterns, but P&G use the systems to model different combinations of detergent based on their underlying molecular structure, while the University of Illinois has been working on HIV research. Audio-identifying app Soundhound is in fact run on an NVIDIA-powered supercomputer.
Compared to the old generation of K20-family supercomputers, those using K40 instead can look forward to a 25- to 40-percent boost in performance.
The first K40 cards will begin to show up on the market in 2014, with ASUS, Dell, IBM, and HP – among others – all on the list to develop and produce their own cards based on the Tesla K40 GPU standard.