io.net ensures GPU performance through continuous benchmarking, automated health checks every 60 seconds, and provider reputation scoring. Underperforming GPUs are automatically flagged, removed from the pool, and replaced—users are refunded if performance falls below specifications.

Performance Monitoring

Automated Checks:
- GPU compute benchmarks (TFLOPS verification)
- Memory bandwidth tests (compare to spec)
- Temperature and throttling detection
- CUDA version compatibility

Frequency: Every 60 seconds per GPU

Provider Reputation

Performance Score:
- Benchmark results vs. expected performance
- Uptime and reliability
- User ratings

Enforcement:
- Providers with <90% performance score: Warning
- Providers with <80% score: Removed from pool
- Refunds issued for substandard performance

Performance Guarantees

Specifications:
- RTX 4090: 24GB VRAM, 82.6 TFLOPS (FP16)
- A100 (80GB): 80GB VRAM, 312 TFLOPS (FP16)
- H100 (80GB): 80GB VRAM, 1,979 TFLOPS (FP16)

If GPU underperforms:
- Automatic migration to replacement GPU
- Refund for affected period

Verify Performance

# Run benchmark on your GPU
io exec my-instance -- nvidia-smi --query-gpu=name,memory.total,power.limit --format=csv

# Run CUDA benchmark
io exec my-instance -- /usr/local/cuda/extras/demo_suite/deviceQuery

Performance-guaranteed GPUs — io.net with continuous benchmarking and auto-replacement.