Scaling figures
The SVG charts in this directory are drawn by tools/render_scaling.py from the CSV extracts in data/. The public site does not need the research repository at render time.
The scientific owner is the communication-topology-scaling study in the research repository. These pages were checked against that repository’s master at ac36f14058e73aad4997799de5229a211c6032e0. The website copy is a frozen extract, not the campaign record.
Machine
LUMI-G, AMD MI250X. One MPI rank per graphics compute die (GCD). Eight GCDs are one node. Node counts in the charts are gcds / 8.
Charts
The homepage shows fd-orders.svg and fft-envelope.svg. The strong-scaling pair is the 171.3× number on that page. fd-weak.svg and heat-strong.svg stay in this directory.
fd-weak.svg and fd-orders.svg are weak scaling of heat3d_fd_hip. The local interior is a constant 256 cubed cells per GCD. dt is 0.01 and I/O is off during the timed steps. Plotted efficiency is that stencil order’s own 16-GCD median step time divided by the later median. The 8-GCD row is one node. It is the other baseline, stored as weak_eff, and it is not the plotted 100% line.
Source table: articles/communication-topology-scaling/data/heat3d_fd_order_campaign_256.csv, blob d885abefe0c7226bb5ef8dd99735aad475de1907. The timed binary is OpenPFC 03e9f67bb42b2570afb074fe653abaa1c68084f1.
heat-strong.svg is fixed-problem FD2 on a global 4096 cubed grid, 30 steps, warmup 5, dt 0.01. The ideal line is the 4-node median times 4 / nodes. Speedup is the 4-node median divided by the later median. Fraction of ideal is that speedup divided by nodes / 4.
Source table: presentations/openpfc-project-update-2026-09-29/data/heat3d-strong-4096.csv, blob dbf446b64b22f00fae02b1de8e6b776080984ee2. The evidence note that admits the ladder is articles/communication-topology-scaling/evidence/2026-10-03-envelope-observations.md. The timed binary is OpenPFC d0f6bf8b397fdfbc9961f55cdf3e81ccae361fb8.
fft-envelope.svg is the 768 cubed local box per GCD. At each GCD count from 16 upward, the point is the fastest of the four measured HeFFTe reshape protocols, divided by the fastest 16-GCD protocol. The 512 cubed family is not in this chart. The curve stops at 4096 GCDs, which is 512 nodes. No 1024-node timing was admitted.
Source table: articles/communication-topology-scaling/data/heffte_protocol_tournament_scaling.csv, blob 37279afa8f6a58650e4f4eed68dd84013910e273. The frozen binary is OpenPFC 8c0fb7637095e84de5a3ae3e4eb4ac779dcff165.
Not drawn
tungsten_hip_strong_1280.csv was read and left out. The ladder stops at 128 GCDs, and the 96-GCD rung does not divide the grid. It is not a second strong-scaling claim.
Headline rounding is one decimal place, from the wall-time ratios, as recorded in manifest.json. Regenerate with:
python3 tools/render_scaling.py
python3 tools/render_scaling.py --check