DownUnder GeoSolutions are giving a talk about how they use thousands of Intel Xeon Phi processor instead of GPU’s in their high performance computing setup.
Although general-purpose GPUs have all of the mindshare, we’ve deployed thousands of the Intel Xeon Phi processors, putting us in stark contrast to most of the commercial HPC world.
If you’ve never heard of them, you’re not alone. Xeon Phi is the product name for Intel’s “Many Integrated Core” architecture, which combines 60+ x86 cores, AVX-512 vector instructions, and high-performance memory into a single package.
This talk discusses:
- Why we chose the Phi, when almost everyone else chose GPUs
- How and why we build the clusters that we do, including a novel solution to the cooling challenge of putting 5 teraflops in each 1 RU
- Programming techniques that resulted in better performance than Intel’s own engineers thought possible, and
- A look at our new 14 petaflop machine, based on the next generation of Phi
- DUG (DownUnder GeoSolutions)