Maybe Paolo insists on his point of view simply because he is actually right.
I know nothing about DOCSIS, but here we are talking about modems and speeds in the order of 10Mbps or some multiples of 10 Mbps, right? When we say performance "impact", we are talking about a performance degradation that can be observed and in this case from the perspective of a typical user and not from the perspective of a single device in the path from source to destination.
Let's say you have $1000 and while you are not looking I take $1 away from you and immediately afterwards I put back $1000. You end up with $1999. Even if you actually saw me taking the $1 away, are you still thinking about that dollar? (ok, maybe you are, but, just to be clear, I'm not giving it back no matter what). If we improve the experience of a typical user by say doubling the bandwidth and we worsen some other negligible parameter, user will only observe an overall improvement.
In networking, latency typically consists of transmission, propagation and queueing/forwarding delay. In many cases, the dominating latency is the propagation delay, which depends on physical distance between source and destination and, in the best case, is determined by the speed of light in optical fiber. Propagation delay is usually in the order of milliseconds (roughly 100ms RTT for a transatlantic link). There is not much that can be done about this component of latency, except perhaps make some links have less length. We can't change the speed of light yet. When we are talking about speeds of 8 Mbps, it means we can transmit a 64-byte frame in 64 usec and larger frames will need more usecs. When you increase the transmission speed, say double it, you can cut the usecs in half. So, how many instructions are needed to forward a packet and how many usecs or msecs do they take? And what's the difference in that delay if we add some extra headers? Because those numbers are negligible compared to other delays, in low end devices you can have inexpensive types of DRAM in the forwarding path.
Since processor types were mentioned, Amdhal's Law/law of diminishing returns came to mind. So, if you wanted to improve the latency observed by a typical network user, would you reduce the number of local hops and number of headers (BTW: "local" means that the propagation delay is negligible in this case) or would you increase transmission speed?
Also, if user downloads a lot of stuff, then the pipe between source and destination becomes full and the delay observed by the user tends to depend on transmission speeds/available bandwidth only.
If we wanted to built a device that can forward packets at multiple Gbps of wire speed, then we would be discussing about high performance internal interconnects, SRAMs, CAMs, TCAMs, ASICs, etc and count instruction clock cycles. In practice, if the devices in a path from source to destination (routers, switches, optical equipment) can sustain the rates they are supposed to serve, then the dominating factor in the latency a typical user experiences is the physical distance from source to destination. If you actually do the math to calculate the delay from say South East Europe to London, you will see that the RTT measured by the ping program can be very close to the theoretical delay a theoretical direct fiber between endpoints would have despite of the numerous equipment between source and destination and the fact that fiber in practice doesn't follow a straight line between the endpoints. Of course, the transmission speeds/available bandwidth also matter.
In any case, a parameter that seems negligible in one context might gain significance in another. If we were discussing high performance networking and computing for specialized environments, then we would try to squeeze any delay component we can put our hands on. On the other hand, when we try to help a customer in their decision process to built a simple network we discuss the factors that really matter to them based on their needs and budget. I can buy the the best and most expensive laptop and DSL connection in the market, but what's the point if I only browse cisco.com ?
