Cloud Olympics — Catalog
The hero artifacts behind the scores · one tile per event · provenance-gated ← the board
In actual Olympics there are photos and videos of the accomplishments. This is that:
the actual output of each scored run — images, a 3-D viewer, a transcript, generated text —
never just a number about it. Every hero names its run; every figure here is in
docs/measured.md.
A tile with no hero yet says so, with the run that is waiting to be published.
Gaussian splatting
on the board →
See also: the Venus de Milo — a real captured 3-D Gaussian splat (Louvre, via superspl.at) at gallery fidelity, the same technique this event scores a 2-D optimisation core of. The Aphrodite is one example; this fit is ours, end to end.
AzureTesla T4 · 20,000 anisotropic Gaussians · 512²
249,353Gaussian-steps/s$0.0028
GCPNVIDIA L4, Cloud Run europe-west4
212,843Gaussian-steps/s$0.0037
Azure · splat-gpu ·
GCP · splat-gpu (europe-west4) ·
splat-gpu-kluf6fo · 2026-09-02 · ff09f71GCP · splat-gpu (europe-west4) ·
splat-gpu-49vxc · 2026-09-02 · ff09f71FluidX3D lattice-Boltzmann CFD
on the board →
GCPNVIDIA L4, Cloud Run europe-west4 · D3Q19 FP16S
2,834MLUPs/s$0.199
AzureTesla T4, Container Apps
2,787MLUPs/s$0.285
Azure · fluidx3d ·
GCP · fluidx3d (europe-west4) ·
fluidx3d-doj9kj5 · 2026-09-02 · 706dd79 (hero pending)GCP · fluidx3d (europe-west4) ·
fluidx3d-zvmnf · 2026-09-02 · 706dd79 (hero pending)Blender Open Data
on the board →
Same baked Blender 4.5.0 scene bundles the benchmark runs — the benchmark scores samples/min and never rendered a picture; this does.
GCPNVIDIA L4, Cloud Run europe-west4 · official launcher, 4.5.0
2,559.1samples/min$0.025
AzureTesla T4, Container Apps · same launcher and scenes
1,085.5samples/min$0.021
rates · blender-bench · Azure
stills · blender-stills (GCP L4, europe-west4) ·
blender-bench-nftsn14 · GCP blender-bench-r2gpz · 2026-09-02stills · blender-stills (GCP L4, europe-west4) ·
blender-stills-9pnzg · 2026-09-02 · hero built by the Visuals LeadVision — structure from motion
on the board →Drag to orbit, scroll to zoom — two tabs: the 1fps sequence (all 22 fragments) and the 4096-word loop-detection run (its largest component only; the job archives only the largest component's points).
GCPCOLMAP · 8 vCPU · 510 of 1,600 registered, 22 disconnected components
26.131.9% registered$0.59
loop detection4096-word vocab tree · 547 of 1,600 registered, 19 components
3.734.2% registered$3.52
GCP · reconstruct ·
GCP · reconstruct (loop detection) ·
reconstruct-9clrb · 2026-08-29 · 944773bGCP · reconstruct (loop detection) ·
reconstruct-jpvzg · 2026-09-02 · 944773b · hero built by the Viewers LeadFrame rendering
on the board →
GCPNVIDIA L4, Cloud Run europe-west4 · identical scene, samples, size
95.0frames/min$1.21
AzureTesla T4 · Cycles · 128 samples · 480² · no denoise
65.8frames/min$3.70
rates · render-gpu · Azure
frames · render-gpu (GCP L4, europe-west4) ·
render-gpu-i0w27lh (65.8/min) · GCP render-gpu-hdfj2 (95.0/min)frames · render-gpu (GCP L4, europe-west4) ·
render-gpu-g6c8w · 2026-09-02, 90.99/min this rerun · hero built by the Visuals LeadImage generation
on the board →
GCPNVIDIA L4, Cloud Run europe-west4 · SDXL-Turbo · 4 steps · 512×512 · 24 images
66.0images/min$1.68
AzureTesla T4 · SDXL-Turbo · 4 steps · 512×512 · 24 images
61.1images/min$1.83
Azure · imagegen-gpu ·
GCP · imagegen-gpu (europe-west4) ·
imagegen-gpu-mbtpb3s · 2026-08-29 · commit ae44d36GCP · imagegen-gpu (europe-west4) ·
imagegen-gpu-7xckj · 2026-09-02 · commit ae44d36 · hero built by the Viewers LeadSelf-hosted LLM inference
on the board →
AzureTesla T4 · Qwen2.5-1.5B · fp16
36.1tok/s$52.4
GCPNVIDIA L4, Cloud Run europe-west4 · same model, fp16
29.7tok/s$41.3
Azure · llm-gpu ·
GCP · llm-gpu (europe-west4) ·
llm-gpu-0dn96x9 · 2026-08-29 · commit bf8a95fGCP · llm-gpu (europe-west4) ·
llm-gpu-jnktw · 2026-09-02 · commit bf8a95f · hero built by the Viewers LeadPhotonics cavity sweep
on the board →
GCPLegume GME · 2,000 geometries · 0 failed
197at once, 514/min$0.51
GCP · photonics (CPU, 200 tasks) ·
photonics-fldzd · 2026-08-29 · hero built by the Visuals LeadSpeech to text
on the board →
AzureTesla T4 · float16
14.3min of audio/min—
GCP · source gs://cloudup-artifacts-1003714208400/asr-mission-v2/transcript.txt · reviewed 2026-09-02 by the Viewers Lead · 773 words, withheld (personal/client content, per specs/009) · docs/measured.md: “Speech is NOT costed”
Cold start
on the board →
AWSEC2 t3.micro
2.9per minute≈$0
AzureContainer Apps
2.5per minute≈$0
multi-cloud · cold-start probes · docs/measured.md “cold start, 2026-08-29” · hero built by the Visuals Lead
Storage throughput
on the board →
AWSS3 · 256 MiB · best of 2, both orders
43.0MB/s down2 runs
GCPGCS · identical object and method
27.2MB/s down2 runs
multi-cloud · bin/storage-bench · docs/measured.md “storage, 2026-08-29” · hero built by the Visuals Lead
Cost of idle
on the board →
TIEAWS0 instances running
$0true zero0
TIEAzure3 jobs, all scale-to-zero
$0true zero0
multi-cloud · config read-back, no single execution (hero pending)
Embeddings
on the board →
AzureTesla T4 · MiniLM-L6 · 602 chunks
59,513chunks/min$0.071
GCPNVIDIA L4, Cloud Run europe-west4 · same model, 602 chunks
19,799chunks/min$0.029
Azure · embed-gpu ·
GCP · embed-gpu (europe-west4) ·
embed-gpu-xid2d67 · 2026-08-29 (hero pending)GCP · embed-gpu (europe-west4) ·
embed-gpu-6s7fh · 2026-09-02 (hero pending)Across events — GPU speedup, four ways
events: imagegen · llm · embed · render · sourced from each event's own RESULT payload
already in docs/measured.md (imagegen-gpu-*, llm-gpu-*, embed-gpu-*, render-gpu-* and their CPU counterparts)
· hero built 2026-09-02 by the Visuals Lead · not one of the 13 board events, so not counted as a tile
10 of 13 heroes published: splat (both clouds, plus the Venus de Milo bonus viewer), blender,
imagegen, render, vision, llm, photonics, cold-start, storage — speech is published too, as an honest
withholding decision, not an excerpt (personal/client content, per specs/009). Still placeholders: fluid,
embed, idle.
bin/catalog-verify refuses any hero whose execution id is not in docs/measured.md,
any hero URL that does not answer 200, or a tile count that does not match the board strip's 13 events.