Kepler is here. And it’s fantastic!
The news came out today that the first Kepler GPU, the GeForce GTX 680, has been launched. A single GPU has 1,536 CUDA Cores. This means that those high-end workstations with 8 PCIe slots will be able to pack 12,288 CUDA cores into a single computer. That’s some serious computational power.
Current high-end Fermi cards have 512 cores, so this new Kepler architecture boasts 3X the number of computation cores.
Normally we focus on the higher-end Tesla products because those more aptly fit the needs of our science, engineering, and financial computing readers. But we are excited nonetheless by this GeForce GPU. It is a major step forward in GPU technology. And this GeForce card portends great stuff for the technical-computing-focused Tesla revamp due out later this year.
As @codedivine points out on Twitter, “The only disappointment is that FP64 rate is very poor at 1/24 of FP32.” So we look forward to the Tesla parts boosting the relative double-precision allocation for compute purposes.
A whitepaper describing the new GeForce card is available for download. The default version comes with 2GB of memory. Also, a big emphasis was placed on making it lower power. These specs, combined with the big leap forward in computation cores, makes this a big day indeed for GPU computing.
Comments 3
Because the TESLA cards are terribly overpriced and GTX cards has significantly crippled double precision performance, there is a serious reason to focus on AMD graphics cards more intensively.
So far is jacket and libjacket support for AMD GPUs incomparable with NVIDIA GPUs based on CUDA, but I think that jacket customers will be more then happy for better AMD GPUs support.
Kepler has signifficanlty crippled double precision as well as compute capabilities. In fact, it’s worse than GTX 580 for compute. Plus, there are not motherboards in the nature with 8 by x16 PCIe slots.
There are servers like vdactr8 from renderstream, there extenders that have 8 slots (you don’t need electrical x16 as x8 is usually enough) though one of them You have to use for comunication card to Your host computer. After all we have srx from evga that has seven of those..the worst problem is not the computer, but the software..