Tesla K10 and K20 announced
Last modified on Wednesday, 16 May 2012 07:42
Nvidia used the GPU Technology Conference, currently taking place in San Jose, California, to announce two new Tesla GPUs based on the Kepler artchitecture, the Tesla K10 and K20.
When compared to previous Tesla generation products based on Fermi architecture, the new Kepler based Tesla's will bring SMX streaming multiprocessors, that provide up to three more performance per Watt when compared to Fermi, Dynamic Parallelism feature that enables GPU threads to dynamically spawn new threads allowing the GPU to adapt dynamically to the data as well as a new Hyper-Q feature that enables multiple CPU cores to simultaneosly use the CUDA architecture cores on a single GPU. The Hyper-Q dramatically increases GPU utilization, slashing CPU idle times and advancing programmability thus making it ideal for cluster applications that use MPI, according to Nvidia.
The newely introduced Tesla K10 is able to produce peak double precision floating point performance of up to 0.19 teraflops, while single precision floating point performance hoovers at impressive 4.58 teraflops. In case you are wondering, the Tesla K10 is based on two GK104 GPUs for a total of 3072 CUDA cores, 8GB of GDDR5 memory that adds to a 320GB/s of memory bandwidth. The Tesla K10 is for servers only and is aimed at seismic, image, siginal processing and video analytics computing applications.
Unlike the Tesla K10, which was at least detailed to some extent, the Tesla K20's specs are still unknown. It is known that this one will be based on the GK110 GPU. This one is aimed at much more intensive computing applictions like computational chemistry and physics, data analytics, weather modeling and is aimed at both servers and workstations systems. Unlike the Tesla K10 which only features SMX as architecture feature, the Tesla K20 will also pack the mentioned dynamic parallelism and Hyper-Q features.
The dual-GK104 Tesla K10 GPU will be available later this month, while the big-dog K20 is scheduled for Q4 2012.