Why Did Hardware QoS Return to Datacenters After Linux Routers Abandoned It?

Hardware QoS used to be everywhere. Early consumer routers relied on switch-chip schedulers, hardware queues and priority systems to keep networks responsive under load.

Then Linux router firmware — especially in the OpenWrt ecosystem — moved toward software queue management (SQM). Tools like fq_codel and CAKE promised fairness and low latency directly in the CPU.

But modern datacenters running 25, 100 and even 400 Gbps networking rediscovered something important:

Software QoS does not scale indefinitely.

Hardware Scheduling Returns

Modern datacenter switches and NICs perform traffic scheduling directly in hardware using:

• multiple hardware queues
• weighted round robin schedulers
• traffic classes
• priority flow control

These mechanisms allow predictable latency at massive throughput without burning CPU cycles.

Why Software QoS Struggles at Scale

Software queue management works extremely well at home broadband speeds.

But when throughput reaches tens or hundreds of gigabits:

• packet classification becomes expensive
• memory bandwidth becomes the bottleneck
• latency increases due to CPU queueing

This is why hyperscale networks increasingly rely on hardware scheduling in switch ASICs and NICs.

Older Router SoCs Face the Same Reality

Interestingly the same problem now appears in older consumer routers.

Many devices built between 2013-2018 use SoCs with:

• single-core or dual-core CPUs
• 600-900 MHz clock speeds
• limited memory bandwidth

Examples include platforms like:

• AR9331 / AR934x
• MT7620 / MT7628
• early MT7621 devices

These routers can still run modern Linux kernels, but the networking stack has become significantly more complex.

Features like SQM, VPNs, firewall rules and container-style services quickly overwhelm the CPU — even on 100 Mbps or 1 Gbps links.

Rediscovering Hardware QoS

Many of these SoCs already include hardware switching fabrics capable of:

• multiple hardware queues
• DSCP or 802.1p classification
• weighted schedulers

Using these hardware capabilities allows routers to maintain good latency without heavy CPU processing.

The mechanisms are simpler than advanced software algorithms like CAKE, but they require almost zero CPU overhead.

Conclusion

Hardware QoS never truly disappeared. It simply moved out of the spotlight.

Datacenters rely on it because software cannot schedule packets at hundreds of gigabits per second.

Older router hardware faces a similar challenge today: modern Linux networking stacks are simply too heavy for small embedded CPUs.

Router operating systems designed with latency-first principles and sane defaults — such as RouterWRT — can extend the life of existing hardware by making better use of the switch silicon that already exists inside these devices.

Sometimes the best networking innovation is not a new algorithm — but using the hardware the way it was originally designed.