Nowadays GPUs are a poor choice for this. GPUs have fairly high latency and are designed to be optimal for embarrassingly parallel computations. Sure you can do 10gigabits/sec but because you're batching that'll result in your latency being high-ish on average.

In terms of non-CPU hardware for routing FPGA's are still king if you need a bit of flexibility. In theory a rule table can be defined in configs and then flashed onto the FGPA for line-speed packet processing. Even then, given how slow FPGA's are to flash CPUs are still king when general packet processing is needed.

meh, just use plain cpus; see dpdk

https://fast.dpdk.org/doc/perf/DPDK_19_11_Intel_NIC_performa...

This is 10 years old! The GPU the benchmarks run on has 480 cores, and modern GPUs have >5000, and the available memory has gone from 1.5GB to 32GB.

I'd love to see these numbers run with a modern GPU.

Interesting... looks like it's effective for routing minimum-sized packets. As an aside, this is the epitome of good web design. Fast, clear, proper line-widths... I'm in love :)