Intel xeon phi

User discussions
1

GROMACS version: 2020.3
GROMACS modification: No
Hello dear everyone!
I want to buy xeon phi for calculating free energy from Gromacs. Is this a good idea?
I just bought nvidia tesla not so long ago. In normal calculations, everything is fine,> 100 ns per day for 10,000 atoms, but for free energy <30 ns per day. As it turned out, Gromacs is not able to count on gpu PME for calculating free energy.
Can xeon phi calculate nb, pme and so on?

3

I want to buy xeon phi for calculating free energy from Gromacs. Is this a good idea?

No, modern CPUs are faster, e.g. a dual AMD EPYC server can have up to 64 cores and will provide better performance. In addition, Xeon Phi is a deprecated architecture.

In normal calculations, everything is fine,> 100 ns per day for 10,000 atoms, but for free energy <30 ns per day. As it turned out, Gromacs is not able to count on gpu PME for calculating free energy.

PME on GPUs with perturbed charges will be supported in the upcoming 2021 release.

We will soon have the first beta of 2021 release that you could try out and provide feedback on.

4

Thanks for the news, I’m looking forward to the beta!
I did not mention it above, but I have a workstation, I bought and assembled everything myself and an epyc for 64 cores is expensive for me. Now from the secondary market you can buy phi 7120 30 times cheaper.
I’m also a fan of computer hardware, so I’ll buy phi anyway, just now or later.
Found one caveat. Phi has three modes, native, MPI and coprocessor. Native is not at all interested, MPI seems to me to be slow due to PCIe 2.0 in 7120, but the coprocessor is already more interesting, but this mode can only be used if the cpu is xeon e5. Question. Should it be xeon on socket 3647 or whatever?
And are there any other nuances?

5

Coprocessor mode isn’t supported (we had a prototype but it didn’t give a speedup and therefore was never merged).

6

I had chance to work on Intel Xeon Phi to scale our in-house code. Xeon Phi is one of the biggest failures from Intel, even Intel had to depreciated Phi architecture. Those three modes work on pen-paper, but actual code has lots of branching of (if/else), can’t take advantage of wide vector register.