Submission + - Nvidia launch the era of GPU computing
imann writes: Regarding http://www.nvidia.com/object/tesla_computing_solut ions.html, nvidia open a new era in the high performance computing.
The new brand called Tesla contains a GPU computing processor, a workstation and a 1U server. The workstation owns 2 GPU and the server 4.
NVidia provides also a software environment called CUDA made of a C-compiler, a FFT and a BLAS library and some sample codes for various computing needs (matrix, imaging, etc..). Some optional modules are available for some specific computing code like monte-carlo, Black-Scholes, binomial operations, matlab. Some like matlab sounds to be free whereas you need to buy the others.
What's sounds very impressive in such approach is the power you can expect from this "GPU computing". Regarding Nvidia's spec, the GPU offer 128-processor computing core connected to 1.5 GB of GDDR3 memory that deliver 76.8 GB/sec due to a very large bus (384-bit). This configuration provides (still regarding nvidia's spec) a 500 gigaflops peak... This surely be achieved only with a massively distributed program BUT if scientists are able to reach this performance it looks incredible... Just imagine a P4 or a cell processor can only reach 14Gflops (http://de.wikipedia.org/wiki/Floating_Point_Opera tions_Per_Second). If now you realize that a 1U Telsa server offer 4GPU, it could theoretically reach 2 TeraFlops...
All this stuff seems to start a new era for HPC computing. Using the GPU computing (with some rewritten code) will surely open new doors to scientist to reach real-time computing or to faster cpu-intensive application like genomic.
The new brand called Tesla contains a GPU computing processor, a workstation and a 1U server. The workstation owns 2 GPU and the server 4.
NVidia provides also a software environment called CUDA made of a C-compiler, a FFT and a BLAS library and some sample codes for various computing needs (matrix, imaging, etc..). Some optional modules are available for some specific computing code like monte-carlo, Black-Scholes, binomial operations, matlab. Some like matlab sounds to be free whereas you need to buy the others.
What's sounds very impressive in such approach is the power you can expect from this "GPU computing". Regarding Nvidia's spec, the GPU offer 128-processor computing core connected to 1.5 GB of GDDR3 memory that deliver 76.8 GB/sec due to a very large bus (384-bit). This configuration provides (still regarding nvidia's spec) a 500 gigaflops peak... This surely be achieved only with a massively distributed program BUT if scientists are able to reach this performance it looks incredible... Just imagine a P4 or a cell processor can only reach 14Gflops (http://de.wikipedia.org/wiki/Floating_Point_Oper
All this stuff seems to start a new era for HPC computing. Using the GPU computing (with some rewritten code) will surely open new doors to scientist to reach real-time computing or to faster cpu-intensive application like genomic.