📊 Performance
Estimateurs de durée de vie du stockage, de capacité de serveur web et d'inférence ML
Estimez la durée de vie d'une carte SD ou SSD selon le volume d'écriture quotidien et la classification d'endurance
Ouvrir l'outil →Calculez le nombre maximal de requêtes par seconde à partir du nombre de workers et du temps de réponse
Ouvrir l'outil →Estimez la latence d'inférence et les FPS pour les modèles ML sur Raspberry Pi et appareils en périphérie
Ouvrir l'outil →Capacity planning estimates for storage lifetime, web server headroom and machine learning inference throughput — the back-of-envelope numbers you want before committing to hardware.
These are deliberately rough. Their job is to tell you whether a plan is off by an order of magnitude, which is the error that actually costs money; they are not a substitute for benchmarking the real workload.
Which tool do I want?
| If you… | Use |
|---|---|
| You are estimating SSD or drive lifetime | Durée de vie du stockage |
| You want to know how much traffic a server handles | Capacité serveur web |
| You are sizing hardware for model inference | Estimateur d'inférence ML |
Averages hide the peak that actually breaks things
A server sized for average traffic falls over, because traffic is not distributed evenly. A site averaging 100 requests a second frequently sees 500 in a burst, and it is the burst that exhausts the connection pool. A common planning rule is to size for peak at three to five times average, and to know your own ratio if you can measure it.
Write amplification is the equivalent trap in storage. An SSD does not write exactly what you send: the controller rewrites blocks during garbage collection, so an application writing 1 TB can put 2–4 TB through the flash. Divide an endurance rating by application writes and you overestimate lifetime, sometimes by several times. Databases with small random writes are the worst case.
For inference, memory bandwidth usually binds before compute does. A model that fits in RAM but streams weights from slow memory each token is limited by that path, not by the processor — which is why a device with modest compute and fast unified memory can outrun one with the reverse.