Cloud Olympics
GCP · AWS · Azure · since 27 Aug 2026 · updated 2 Sep 2026catalog →
photonics
speech
start
splat
fluid
blender
imagegen
render
vision
embed
llm
storage
idle
Gaussian-steps, per secondCOMPLETEDhero →
entrantper second$ / 1k
AzureTesla T4 · 20,000 anisotropic Gaussians · 512² · PSNR 39.45 dB in 120 s
249,35312.5 steps/s$0.0028◀
GCPNVIDIA L4, Cloud Run europe-west4 · same 20,000 Gaussians · PSNR 39.12 dB in 141 s
212,84310.6 steps/s$0.0037◀
lattice updates, per secondCOMPLETEDhero →
entrantmillion / s$ / T
GCPNVIDIA L4, Cloud Run europe-west4 · FluidX3D D3Q19 FP16S · 99.3% of the vendor's L4 figure · 219 GB/s in 63 s
2,834MLUPs/s$0.199◀
AzureTesla T4, Container Apps · 97.2% of the vendor's T4 figure · 216 GB/s in 64 s
2,787MLUPs/s$0.285◀
AWSTesla T4, EC2 g4dn.xlarge us-east-2 · same image digest · 97.3% of the vendor's T4 figure · 216 GB/s in 65 s
2,793MLUPs/s$0.362◀
Blender Open Data, samples per minuteCOMPLETEDhero →
entrantper minute$ / scene
GCPNVIDIA L4, Cloud Run europe-west4 · official launcher, Blender 4.5.0 CUDA · monster 1,313.5 · junkshop 664.6 · classroom 581.0 · 158 s · not submitted
2,559.13 scenes$0.025◀
AzureTesla T4, Container Apps · same launcher and scenes · monster 547.2 · junkshop 295.5 · classroom 242.8 · 136 s
1,085.53 scenes$0.021◀
frames registered, per minuteCOMPLETEDhero →
entrantper minute$ / 1k
GCPCOLMAP · 8 vCPU · 510 of 1,600 registered, in 22 disconnected components
26.131.9% registered$0.59◀
loop detection4096-word vocab tree · 547 of 1,600 registered, 19 components · 2 h 27 m · does not join the fragments
3.734.2% registered$3.52◀
exhaustive matchersame clip, 3,200 frames — 5.1M pairs, timed out at 60m
0no result—◀
frames rendered, per minuteCOMPLETEDhero →
entrantper minute$ / 1k
GCPNVIDIA L4, Cloud Run europe-west4 · identical scene, samples and size
95.00.63 s each$1.21◀
AzureTesla T4 · Cycles · 128 samples · 480² · no denoise
65.80.91 s each$3.70◀
GCP8 vCPU · identical scene, samples and size
13.14.57 s each$3.98◀
AWSFree Plan account — no GPU types until upgraded (A-060)
——◀
images generated, per minuteCOMPLETEDhero →
entrantper minute$ / 1k
GCPNVIDIA L4, Cloud Run europe-west4 · SDXL-Turbo · 4 steps · 512×512 · 24 images · cheaper per image
66.00.91 s each$1.68◀
AzureTesla T4 · SDXL-Turbo · 4 steps · 512×512 · 24 images
61.10.98 s each$1.83◀
GCP8 vCPU · same image · 34× slower
1.832.5 s each$22.72◀
tokens decoded, per secondCOMPLETEDhero →
entrantper second$ / 1M
AzureTesla T4 · Qwen2.5-1.5B · fp16 · 3 runs, all 36.1
36.1256 new tokens$52.4◀
GCPNVIDIA L4, Cloud Run europe-west4 · same model, fp16 · 3 runs 29.4–29.7 · cheaper per token
29.7256 new tokens$41.3◀
GCP8 vCPU · same model, CPU
3.510.3× slower$170.9◀
AWSFree Plan account — GPU instance types refused until upgraded (A-060)
——◀
cavity geometries solved at onceCOMPLETEDhero →
entrantat once$ / 1k
GCPLegume GME · 2,000 geometries · 0 failed
197514 per min$0.51◀
Azureno task index in Container Apps — a partitioned sweep cannot be expressed
—capability gap—◀
Ansys FDTD29 lab licences, 9 per GPU job
3licence ceiling—◀
min of audio, per minuteCOMPLETEDhero →
entrantper minutecost
AzureTesla T4 · float16
14.35:16 elapsed—◀
GCP80 vCPU, 20 tasks · int8
5.314:18 elapsed—◀
AWS16 vCPU, 8 boxes — unverifiable, RunInstances denied
3.223:36 elapsed$0.30◀
machines ready, per minuteCOMPLETEDhero →
entrantper minutecost
AWSEC2 t3.micro — unverifiable, RunInstances denied
2.920.6 s each≈$0◀
AzureContainer Apps · 24-29 s observed
2.532.8 s each≈$0◀
GCPCloud Run job · 68-114 s, 1.7× spread
0.968.0 s each≈$0◀
MB down, per secondCOMPLETEDhero →
entrantper secondup
AWSS3 · 256 MiB · best of 2, both orders
43.05.95 s2 runs◀
GCPGCS · identical object and method
27.29.40 s2 runs◀
AzureBlob · identical object and method
16.115.87 s2 runs◀
billed while doing nothingCOMPLETEDhero →
entrantidle costheld
TIEAWS0 instances running
$0true zero0◀
TIEAzure3 jobs, all scale-to-zero
$0true zero0◀
GCP1 service pinned warm
~$7per month1◀
chunks embedded, per minuteCOMPLETEDhero →
entrantper minute$ / 1k
AzureTesla T4 · MiniLM-L6 · 602 chunks
59,5130.6 s encode$0.071◀
GCPNVIDIA L4, Cloud Run europe-west4 · same model, 602 chunks · cheaper per chunk
19,7991.8 s encode$0.029◀
GCP8 vCPU · Free Plan account — no GPU types until upgraded (A-060)
8,7794.1 s encode$0.030◀
Azuresame box, CPU image — the 7.3× the GPU buys
8,1944.4 s encode$0.028◀
AWSnot entered
——◀
Route by sport
GPU work — only entrant with oneAzure
anything under a minuteAWS
no rule yet — 0 wins in 2—
What is blocking each sport
speechscored, all three lanes — but NOT costed: no surviving execution matches the published elapsed times, so the old \$0.32 could not be sourcedDONE
startscored, all three lanesDONE
storageAzure leg not entered yetDONE
idlescored, all three lanesDONE
visionA learned matcher (SuperPoint+LightGlue, L4, 09-03) took registration from 31.9% to 59.1% at the same cost per registered frame and 8.7× faster than loop detection — but components went 22 → 39 and the largest shrank 122 → 75 frames. The matcher was worth ~2×; it cannot join a scene the camera never tied together. Next lever is capture, not another matcher. Closed as an algorithm questionOURS
All three GPU lanes are open
AWS
OPEN 2026-09-03 — the operator upgraded the account off AWS's Free Plan and a real g4dn.xlarge (Tesla T4) launched, ran FluidX3D from the same image digest as the other two clouds, and scored 2,793 MLUPs/s. The SCP still denies Spot and us-east-1; on-demand in us-east-2 is what the lane usesDONE
GCP
CLEARED 2026-08-31 — trial upgraded. T4×8, L4×16, V100×8, 3000 vCPU; a real T4 instance launched and was deleted. GPUS_ALL_REGIONS is gone from project-info entirelyDONE
Azure
Tesla T4 16GB proven 23 Aug, driver 580.159.04. Serverless, no quota request neededDONE
Third column is list price per unit of work — live vendor pricing APIs, measured durations, before any credit or free tier. Faster is not always cheaper.11 events scored · one full pass costs $3.62 at list price