scope+bench(phonon): config-batching proof-of-mechanism (verdict: park until H100) - #288
Draft
wladerer wants to merge 3 commits into
Draft
scope+bench(phonon): config-batching proof-of-mechanism (verdict: park until H100)#288wladerer wants to merge 3 commits into
wladerer wants to merge 3 commits into
Conversation
Isolates the gating question for phonon config-batching: fold N displacement configs into the leading batch axis (N·nk, nb, npw_max) and time one folded batched Davidson vs N separate solves, with a per-config eigenvalue correctness check. Does NOT run a full config-batched SCF (per-config v_eff/density/Fermi/ mixing) — just measures whether the batched-eigensolve lever is real before committing to that build. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Scope doc (docs/design/phonon-config-batching.md) + probe measurements. Verdict: DO NOT build the config-batched SCF driver yet. Folding is bit-exact (fp64 Δeig 7e-14) but neutral-to-negative on all accessible hardware — CPU 0.46x (BLAS already saturated), RTX 3050 GPU 1.07x (too small to under-fill with a realistic spoke). The GPU-saturation premise is an H100-class claim, untestable on the 3050. Park until H100; SeedPool process-parallelism stays the campaign lever. Fix probe fp32 correctness threshold (2e-4 eV is fp32 noise, not a fold bug). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Scopes phonon config-batching (fold N displacement configs into the plane-wave
batch axis for one
(N·nk)SCF, vs today's serial / SeedPool-process spokes) andmeasures the one lever that gates the whole build before building it.
Deliverables
docs/design/phonon-config-batching.md— full scope: what folds (G-sphere/kinetic/position-free projectors) vs per-config (v_eff/density/Fermi/mixing), the required
changes (per-row
v_effinBatchedHamiltonian, a config-batched SCF driver withper-config freezing), risks, and the go/no-go gate.
benchmarks/phonon_batch/config_batch_probe.py— isolates the eigensolve lever:N separate
davidson_batchedvs 1 folded(N·nk)solve, with a per-configeigenvalue correctness check. Does not run a full config-batched SCF.
Result (asus) — NO-GO on current hardware
The fold is bit-exact (fp64). But it's neutral-to-negative everywhere testable: CPU
BLAS is already saturated at one SCF's
nk; the 3050 is too small to under-fill with arealistic phonon spoke (and OOMs on the larger fold). The GPU-saturation premise is an
H100-class claim, untestable here.
Recommendation: don't build the config-batched SCF driver yet — validate the premise
on an H100 first (one spoke under-fills → fold fills it). Until then SeedPool stays the
campaign lever. The one unique advantage worth revisiting on H100: config-batching is
autograd-compatible, which SeedPool's process-parallelism structurally is not.
Draft — this is a scoping + measurement artifact, not a feature.
🤖 Generated with Claude Code