Nvidia releases final version of CUDA 5.0

Nvidia releases the preview version of GPGPU CUDA 5.0 interface today with the final version. Example, it supports the Dynamic Parallelism, can be generated by a thread on the GPU expiring dynamic new kernel. This reduces the CPU load, ease of programming and allows the execution of more complex algorithms than previously. On the Nvidia GPU Technology Conference had declared in May 2012 that developers could save by the new function and program code: Sun could break down a long 200-300 lines of code to about 30 lines.
Dynamic Parallelism dominate but only Nvidia's upcoming GK110 GPUs. They consist of more than 7 billion transistors, true computing monster and first appear on the supercomputer computing Tesla cards K20. Some of which already count in the Oak Ridge National Laboratory - and there are more and more: the total want to equip the Oak Ridge Labs Titan supercomputer with 14,592 K20 computational cards. The task should be completed in 2013. Consumer graphics cards with GK110 GPUs will be taken next year on the market, rumored earlier in March 2013 for the next GPU Technology Conference. Nvidia wants to first meet the supercomputing needs - where the chips also bring more money. A Tesla K20 will cost 3200 U.S. dollars.

CUDA 5.0 also brings innovations in GPUDirect technique that allows direct Datenaustautsch between the memories of different GPUs. CUDA 5.0 for this to work now has a network connection between graphics cards, which are in different servers (Remote Direct Memory Access / RDMA). Also there are GPU Library Object Linking and Nsight Edition Eclipse. According to Nvidia CUDA 5.0 is available for download starting today.

No comments:

Post a Comment