Intel Xeon Phi is much easier to program than Tesla
Just yesterday Intel announced its cGPUs for massively parallel computing Xeon 5110P Phi , and is beginning to advertise, showing the advantages to rival products ( Tesla K20X Nvidia and S10000 FirePro of AMD).
The parallel computation accelerator 5110P Intel xeon Phi (1.01 TFlops in double precision calculations) pales before the power of Tesla K20X (1.31 TFlops in double precision) and FirePro S10000 (1.48 TFlops in double precision), but before this disadvantage to rivals Intel claims that their product is much easier to program, allowing easily leverage its 60-core x86 (capable of running 240 threads of processing) and 960 shader processors (vector units grouped into 60 16-Wide).
To prove their claims Intel provides an example showing how to adapt common code ( Monte Carlo algorithm for this example), which was implemented in 693 seconds on a computer based on a Dual Xeon configuration (dual socket), which with minimal changes (two additional lines of code) was run in 6.35 seconds on a Xeon Phi.
Do the same for the throttle Nvidia Tesla takes hard work to pass the standard code (C/C + + or Fortran) to CUDA, plus it would be very difficult to re-translate for execution on a standard CPU.
Link: What does it take to code for a Xeon Phi? (SemiAccurate)Tags: 5110P, accelerator, C + +, cGPU, comparison, CUDA, Fortran, openmp, parallel, Phi, programming, server, supercomputer, TACC, Texas Advanced Computing Center, x86, xeon