Packet → NIC → Host CPU → nftables (kernel) → Forward/Drop → Host CPU → NIC → Wire Every packet consumes CPU cycles, limiting throughput, especially at 10 GbE, 25 GbE, or higher.
With kmod-nft-offload + compatible hardware: kmod-nft-offload
In short, it allows certain nftables rules (e.g., forwarding, DNAT, SNAT) to be programmed directly into that supports flow offloading. How It Works Without offload: Packet → NIC → Host CPU → nftables