Szilárd Páll

HPC researcher, molecular dynamics parallel algorithms, core-dev of GROMACS; accelerator geek.

[HGPU group @hgpu]

GROMACS on AMD GPU-Based HPC Platforms: Using SYCL for Performance and Portability

#Physics #MolecularDynamics #HPC #Performance #SYCL #Package

https://t.co/5K4fUhbuR1

[Rossen Apostolov @RossenApostolov]

We are hiring and looking for experts in either Molecular Dynamics (GROMACS, https://t.co/XvoqGOTiDY) or CFD (Neko, https://t.co/2RwktEUOBE) to join our teams at KTH. Deadline for applications is 24 May.

https://t.co/F3T2ZNm2vn

[Szilárd Páll @twilard]

@thedeadline @JeffdotLayton One Tien I was frantically waving my hands in the sink, a conference center staff showed me the trick: move your hand close to the sensor, almost touching it, if soap is not coming out it's probably broken or empty, so move on to the next one. Worked like a charm ;)

[Szilárd Páll @twilard]

@ifndef_define Good idea, it's been a terrible experience waking up with cracked lips and a nosebleed. However, hotel rooms tend to be huge in the USA, I suspect that portable humidifiers won't be able to get typical hotel rooms to a comfortable humidity level.

[Szilárd Páll @twilard]

@eriklindahl @FrankNoeBerlin @JavanainenM @LindorffLarsen @CompBioPhys @CG_Martini @ProfSyK @attecscfi @nvidia To be fair, they use 14 nm FPGAs vs 7 nm A100s GPUs so there's room for perf/W & perf density) improvements.
Also, the custom 3D-torus network is a major advantage vs current HPC interconnects which are not great for strong scaling MD, so there is potential if perf/$ makes sense

[Szilárd Páll @twilard]

@eriklindahl @FrankNoeBerlin @JavanainenM @LindorffLarsen @CompBioPhys @CG_Martini @ProfSyK @attecscfi @nvidia While the results are decent, nowhere near 20x perf / 200x energy advatage. At peak, the perf vs GROMACS on 16/32-nodes of JUWELS-Booster is ~3/4x, energy improvements is max ~4-5x.

The paper is also very scarce on algorithmic details. I'd love to see parallel efficiency too

[Szilárd Páll @twilard]

@science_dot @HatemLtaief @IXPUG1 2x is surprisingly large, the only reason I could see for that is different core counts or perhaps comparing Rome to Milan-X. Also note that this is NAMD 2 (not the GPU-resident v3) which is often CPU-bound as it streams data off of the GPU for integration on the CPU.

[Kenneth Hoste (mastodon: [email protected]) @kehoste]

Today at 14:30 UTC, Robert McLay (@TACC) will be presenting a retrospective on 15 Years of Lmod, the environment modules tool written in Lua, during the Lmod monthly meeting; more info via https://t.co/YWyk4yRUjW .
Let's make sure his talk is well attended, spread the word!

#HPC

[Szilárd Páll @twilard]

@micheleweiland @walkingrandomly Not always, best time to solution and energy to solution may be achieved by different algorithms.

Also, optimizing code not on the critical path is traditionally seen unimportant, but it can contribute to better energy to solution (e.g. by letting computer units idle longer).

[Szilárd Páll @twilard]

@walkingrandomly Yeah, users will likely not care until they're charged for energy consumed rather than core-hours.

[Szilárd Páll @twilard]

@hpcnotes I think it's a matter of how you define "accelerator". By "#CPU" are you referring to processors with homogeneous cores and only intra-core "accelerators". What about heterogeneity across cores on-die or package?

[Szilárd Páll @twilard]

@_zzzoom_ @hpcprogrammer Are you suggesting we should stick to dkms hell unless the entire driver stack is opened up?

[HPC Guru @HPC_Guru]

@nvidiadeveloper Grace Hopper vs x86+Hopper

Please note that these are performance simulations, not actual measured performance

#HPC #AI #ARM #GPU @nvidia

[Szilárd Páll @twilard]

@adrianjhpc @JimCownie @giIIes @gilles I understand. It could certainly help the broader community, and perhaps it might even become useful in informing future Archer procurement.

[Szilárd Páll @twilard]

@adrianjhpc @JimCownie @giIIes @gilles Thanks! What is the main purpose, would it not be useful to try to include all parallelization APIs/prog/ models?
e.g. for GROMACS I see "MPI+OpenMP", but that ignores GPUs (I know Archer2 not a GPU machine). MPI+OpenMP+CUDA/SYCL/(OpenCL) would be more accurate.

[Szilárd Páll @twilard]

@hpcprogrammer In codes/algorithms *designed* for heterogeneity, going from N=2 to N>=2 should not be as challenging as going heterogeneous. Some aspects will be more important, e.g. auto-tuning.
That is if there will be *one* prog. model for everything (not one for each _PU/vendor)

[Szilárd Páll @twilard]

@HPC_Guru @nvidia Will this lead to China ramping up strategic investment in AI hardware as it happened for HPC?