− | *** Models are memory hungry, at least 32 GB of RAM is needed. | + | *** Models are memory hungry, at least 32 GB of RAM is needed. Ideally, 64 GB. |
− | *** GPU servers deliver the best performances for prediction, but are rather expensive. Given the current load, CPU servers are enough. High-performance CPU are however needed. 8 cores at least (the more, the better). | + | *** Ideally, 1 GPU server. Otherwise, high-performance CPU are suitable, 8 cores at least (the more, the better). |
− | *** Disk: speed is only important when loading model from disk, which only occurs at startup. SSDs are fine. 500 GB of space should be ok. | + | *** Disk: speed is only important when loading model from disk, which only occurs at startup. Ideally NVMe, classic SSDs are fine. 500 GB of space should be ok. |