loading globe…
Throughput Bench
Globe Race
GPU
Sensor · GSD
0.3 m — high-res commercial
1 m — NAIP / aerial
10 m — Sentinel-2
30 m — Landsat
Model A
Model B
Speedup
1000×
▶ Start Race
Explore data
List data
Model performance
Throughput vs compute cost — click a family in the legend to filter.
Showing fp32 precision, no torch.compile (eager baseline)
×
GPU
X axis
Parameters (M)
MACs (GFLOPs)
Peak memory (MB)
Y axis
Throughput (img/s)
Coverage (km²/s)
Precision &
torch.compile
speedup
Throughput relative to fp32-eager (baseline = 1.0×) for ResNet, ViT, and EfficientNet families.
Models × GPU architectures
Best throughput (img/s) per model on each GPU · click a column header to sort
×
?
About
Don't show automatically
Got it
—
0
km²
0% of land
—
0
km²
0% of land
00.0 s
1×