Alex Izvorski

> It's actually implemented at a very low level. Not the DRAM itself, but just above it. @peastman I had no idea! Okay, so this got me interested in how...

One more toy kernel - update 3 consecutive uint64 "forces" for 1M atoms, where the forces are padded to 32 bytes: ``` add_kernel = cp.RawKernel(r''' extern "C" __global__ void test_add_x3(unsigned...

> and experimented with disabling the single pair lists @dmclark17 Interesting - when you say disable the single pair list, do you mean some forces which should be calculated were...

@peastman @dmclark17 One of the big issues here is the essentially 0% L1 cache hit rate. This is something I really don't get. The memory access pattern should be fairly...

Another issue (and maybe the thing that determines overall speed) is that the L2 hit rate is also surprisingly low. If forces were simply updated sequentially (with atomicAdd), the L2...

@Samgonza `forcefield.registerResidueTemplate(...)` takes a **template** as an argument, not a residue name (res.name in your code). Please have a look at some of the example code in http://docs.openmm.org/7.7.0/userguide/application/05_creating_ffs.html#missing-residue-templates and see...

@peastman Thanks! Could be something to add to the docs? I assumed this, and seemed to be true in simple tests, but didn't have any way to know if it's...

Could be under ForceField createSystem - something like "The constructed system contains one particle for each atom in the topology, and the particles are in the same order as the...

@mwort "typical 5-95% CI" - I believe "95%CI" is 95% wide ie between 2.5% and 97.5%. Is there a reference that points to the other convention?

I have been working with other ase calculators (eg psi4) and can confirm that it is possible to have a net charge for the whole system without having the initial_charges...