Guide
How to estimate bandwidth requirements for AI training
Most teams size GPUs carefully and guess at data transfer. On big clouds that guess shows up as an egress bill after the run. Here is how to estimate it before you launch.
The formula
Bandwidth for a training run is the data moving in plus the data moving out:
data_in = dataset_GB × (streamed ? epochs : 1) + base_weights_GB data_out = checkpoint_GB × checkpoints_pulled + artifact_GB × downloads time_h = (data_in + data_out) × 8 / link_Gbps / 3600
- Checkpoint size: bf16 weights ≈ 2 bytes per parameter (8B ≈ 16 GB, 70B ≈ 140 GB). Optimizer state roughly triples a full-resume checkpoint — keep those on the instance unless you need them.
- Streaming vs caching: streaming from object storage multiplies ingress by epoch count. Copying to local NVMe once is usually faster and cheaper.
- Base weights: on CloudFlux popular models are pre-cached, so they add zero download time.
Worked examples
Big-cloud figures use a typical internet egress list price of $0.09/GB; your provider's tier may differ.
LoRA fine-tune, Llama 3.1 8B
50 GB instruction dataset cached on NVMe, 5 adapter checkpoints pulled, final merged model shipped to 3 serving regions.
In 50 GB · Out 49 GB · ~0 h transfer · big-cloud egress ≈ $4.4 · CloudFlux $0
Full fine-tune, 70B model
1 TB dataset streamed each epoch, 10 full bf16 checkpoints (~140 GB) downloaded for evaluation, final weights pushed to 4 regions.
In 2,000 GB · Out 1,960 GB · ~0.1 h transfer · big-cloud egress ≈ $176.4 · CloudFlux $0
SDXL image model training
300 GB image set cached locally, 8 checkpoints (~7 GB) pulled, 2,000 sample grids (~0.05 GB each) reviewed.
In 300 GB · Out 156 GB · ~0.1 h transfer · big-cloud egress ≈ $14 · CloudFlux $0
The pattern: egress grows with every evaluation checkpoint and every region you deploy to — exactly the habits that make models better. Zero-egress pricing removes the penalty for them.
Bandwidth calculator
- Data in
- 50 GB
- Data out (egress)
- 49 GB
- Typical big-cloud egress bill
- $4.4
- CloudFlux egress bill
- $0
Transfer time at 10 Gbps: ~0 h.
How CloudFlux handles bandwidth
- Bandwidth is never billed — in or out.
- Indie includes 2,000 GB and Enterprise 50,000 GB of egress per job; beyond that, transfers are rate-limited instead of charged.
- 10–100 Gbps networking and pre-cached weights shorten transfer time before the first step.