Parallel Computing for Data Science

6.8. THE INTEL XEON PHI CHIP 177

a large head start in this market, so it is unclear how well the Intel chip

will do in the coming years.

The Xeon Phi features 60 cores, each 4-way hyperthreaded, thus with a

theoretical level parallelism level of 240. This number is on the low end

of NVIDIA chips. But on the other hand, the Intel chip is much easier

to program, as it has much more of a classic multicore design. One can

run OpenMP, MPI and so on. As with GPUs, though, the bandwidth and

latency for data transfers between the CPU and the accelerator chip can

be a major issue.

It must again be kept in mind that while the NVIDIA chips can attain

exceptional ...