Free tool: 30-second GPU benchmark for RunPod pods

[dentity007]

Hey everyone — I built a quick benchmark script during the competition and figured it might save some of you time and money.

The problem: Not all H100s perform the same. Cloud GPUs vary in clock speeds, memory bandwidth, and thermals depending on the specific machine you land on. I burned ~$40 on a pod that was running 30% slower than expected before I figured this out.

The solution: A 30-second shell script that benchmarks your pod before you start training:

• GPU identity + clock speeds
• GEMM throughput (bf16 TFLOPS)
• Memory bandwidth (GB/s)
• Disk I/O + network speed
• Pod location

Quick reference — what "good" looks like on H100 SXM:

Metric	Good	Consider switching
GEMM 4096×4096 bf16	< 0.50ms (275+ TFLOPS)	> 0.70ms
Memory BW	> 2800 GB/s	< 2000 GB/s
Max GPU clock	1980 MHz	< 1800 MHz

Usage:

curl -s https://raw.githubusercontent.com/NathanMaine/runpod-gpu-benchmark/main/pod-test.sh | bash

No dependencies beyond PyTorch (already on every RunPod template).

Repo: https://github.com/NathanMaine/runpod-gpu-benchmark

Hope it helps someone avoid the slow-pod tax. Good luck out there!