AI Training Calculator
Estimate VRAM usage and find a suitable GPU configuration for inference or fine-tuning. All figures are estimates; actual usage varies by framework and implementation.
Parameters
VRAM Breakdown
Inference · FP16 · Batch 1
21.2GB
Model Weights16.0 GB75%
KV Cache0.5 GB3%
Activations3.2 GB15%
Framework Overhead1.5 GB7%
GPU Requirements
View Pricing →A100 40GB
Recommended
VRAM / Card
40 GB
Cards Needed
1x
Total VRAM
40 GB
53%
B200
VRAM / Card
192 GB
Cards Needed
1x
Total VRAM
192 GB
11%
H200
VRAM / Card
141 GB
Cards Needed
1x
Total VRAM
141 GB
15%
H100 NVL
VRAM / Card
94 GB
Cards Needed
1x
Total VRAM
94 GB
23%
A100 80GB
VRAM / Card
80 GB
Cards Needed
1x
Total VRAM
80 GB
27%
H100 SXM5
VRAM / Card
80 GB
Cards Needed
1x
Total VRAM
80 GB
27%
L40S
VRAM / Card
48 GB
Cards Needed
1x
Total VRAM
48 GB
44%
RTX 4090
VRAM / Card
24 GB
Cards Needed
1x
Total VRAM
24 GB
88%
Card counts assume 90% of per-card VRAM is usable, reserving headroom for framework overhead and fragmentation. Multi-GPU setups also depend on interconnect bandwidth.