Serve More Customers With Fewer Servers
per subscriber
total cost
happier customers
(Accelerated QoS)
Think of it as three engines for the same car. They all get subscribers online — but one sips fuel and one drinks it. At ISP scale, that fuel bill is real money — every month.
Software Processing
The traditional way. Maximum flexibility.
- All work done in general-purpose software
- Every feature available, easiest to customize
- Highest hardware use per subscriber
- Best for smaller sites or richest feature needs
Hybrid — Fast-Path Traffic Fairness
The balanced choice. Big saving, no features lost.
- The heavy "traffic fairness & speed-shaping" work is moved to a high-speed fast path
- All address-sharing & security features kept
- Roughly half the hardware per subscriber
- Best all-round balance of cost and capability
Everything in the Fast Path
Maximum density. Lowest cost per subscriber.
- All processing — sharing, security, fairness — in the fast path
- Lowest hardware use & lowest latency
- Most subscribers per server
- Best for large scale & growth
1 · Processing cost per subscriber (lower is cheaper — index, Standard = 100)
How much server "horsepower" each subscriber consumes
Less horsepower per subscriber → more subscribers per server → lower cost.
2 · Servers needed to serve 100,000 subscribers (illustrative)
Fewer servers = lower purchase cost, power, cooling, rack space & maintenance
The single clearest money picture: same subscribers, very different hardware footprint.
3 · Relative 5-year total cost (hardware + power + space + upkeep · Standard = 100)
What the choice is worth over the equipment's lifetime
Driven mostly by server count and the electricity to run & cool them.
4 · Subscriber experience — responsiveness (delay added by the gateway — lower is better)
Faster handling = snappier browsing, gaming & video calls = fewer complaints
Latency is invisible on a bill but very visible to customers — it shows up as churn.
What lower latency means for your customers — games, calls & video
Delay (latency) and uneven delay (jitter) never show up on a bill, but they are exactly what customers feel — and the #1 hidden cause of "my internet is bad" tickets and churn. Acceleration shaves delay off every packet, which lands straight in the apps people care about most:
Responsive, competitive play
- Without: lag spikes, "rubber-banding", high ping, losing winnable fights
- With acceleration: low, steady ping — fast reactions, smooth play
- Lets you sell a premium "gamer / low-latency" plan
Clear, natural calls
- Without: robotic, choppy audio, echo, words cut off, talk-over
- With acceleration: low jitter → crisp, real-time conversation
- Fewer "can you hear me?" complaints
Smooth HD meetings
- Without: frozen frames, audio/video out of sync, "you're frozen", reconnects
- With acceleration: stable, in-sync HD on Zoom / Teams / Meet
- Critical for work-from-home subscribers
Why it's money: better real-time experience → fewer support calls, higher satisfaction, lower churn, and the ability to market premium low-latency / gaming tiers. The latency you can't see on a spreadsheet is exactly what customers leave over.
5 · Room to grow before you buy more hardware (spare capacity at busy hour — higher is better)
Headroom = how many more subscribers you can add for free
More spare capacity means you delay the next hardware purchase — sometimes for years.
Which mode fits which goal?
| Your priority | Standard | Accelerated QoS | Full Acceleration |
|---|---|---|---|
| Lowest cost per subscriber | ○ | ◐ good | ● best |
| Fewest servers / least power | ○ | ◐ | ● best |
| Best subscriber experience | ○ | ◐ | ● best |
| Keep every advanced feature | ● best | ● full | ◐ most |
| Simplest to operate / customize | ● best | ◐ | ◐ |
| Highest scale & growth | ○ | ◐ | ● best |
● strongest · ◐ good · ○ limited
How acceleration actually raises your ceiling
| Server | Processor | Memory | Network & real ceiling | Today (Standard) |
|---|---|---|---|---|
| Server A | Xeon E5-1650 v4 · 6 cores / 12 threads (HT) | 32 GB | 2× 10G → 20 Gbps link cap | ~16 Gbps · ~5,000 users · CPU-limited |
| Server B | Xeon E5-2697A v4 · 16 cores / 32 threads (HT) | 32 GB | 2× 40G XL710 → 56 Gbps (PCIe 3.0 ×8) | ~33 Gbps · ~9,000 users · CPU-bound |
Server A — max subscribers (6c / 12t HT · 2× 10G = 20 Gbps)
Standard is CPU-capped below the link. Acceleration fills the 20G link (→ 6.7k). The link is then the wall — a faster card lifts it.
* A new 2×40G card lifts Server A to the ~56 Gbps PCIe 3.0 ×8 / XL710 ceiling. Its 12 threads spread the receive queues and drive this comfortably for Accelerated QoS; for Full address-sharing at 56 Gbps a CPU refresh adds margin.
Server B — max subscribers (16c / 32t HT · 2× 40G → 56 Gbps PCIe 3.0 ceiling)
Two 40G ports, but PCIe 3.0 ×8 + the XL710 cap throughput at ~56 Gbps. Acceleration fills that ceiling; exceeding it needs a new platform.
Memory: 32 GB is fine for traffic-fairness; large-scale address-sharing (millions of connections) benefits from 64–128 GB — an inexpensive upgrade that extends both servers.
Recommended platforms — to go further
When you want to grow beyond what the current servers and links allow, here are three sensible steps. Acceleration scales with CPU cores × clock speed × network-card line rate; memory sizes the address-sharing tables.
| Platform | Processor | Memory | NIC + PCIe | Usable throughput | Max subscribers @3 Mbps | Best for |
|---|---|---|---|---|---|---|
| ① Keep & Accelerate lowest cost | your current servers | → 64 GB | existing 82599 / XL710 · PCIe 3.0 | 20 – 56 Gbps | ~6,700 – 18,700 | Immediate win, near-zero capex |
| ② 100G Performance | 16-core 3.0 GHz (EPYC 7313P / Xeon Gold 6326) | 128 GB | 2× 100G Intel E810 · PCIe 4.0 | ~100 Gbps | ~33,000 | One node replaces several old ones |
| ③ 100–200G Scale | 32-core (EPYC 7543 / 9354) | 256 GB | 2–4× 100G E810 / ConnectX-6 · PCIe 4.0 | ~150 – 200 Gbps | ~50,000 – 64,000 | Large / Tier-1 growth |
Maximum subscribers per node (Full Acceleration · @ 3 Mbps/user)
More subscribers per node = fewer nodes = lower cost per subscriber. Figures = usable throughput ÷ ~3 Mbps busy-hour (≈ 333 subs/Gbps); depend on traffic mix & feature set.
The smart sequence: start with ① — accelerate the servers you already own (biggest return, lowest spend) — then add ② or ③ only when subscriber growth needs more pipe.
Example: what it saves on 100,000 subscribers
5-year gateway cost — same subscribers, three modes
Illustrative, using an all-in figure of ~$12,000 per gateway node, per year (hardware amortization + power + cooling + rack space + support), and ~3 Mbps/subscriber. Your numbers will differ — the ratio is the point.
| Mode | Nodes for 100k | 5-year cost | You save vs Standard |
|---|---|---|---|
| Standard | ≈ 10 | ≈ $600,000 | — |
| Accelerated QoS | ≈ 5 | ≈ $300,000 | ≈ $300,000 (−50%) |
| Full Acceleration * | ≈ 2–3 | ≈ $150,000 | ≈ $450,000 (−75%) |
That is up to ~$450,000 saved over five years on a 100k-subscriber footprint — plus a faster service that reduces churn. Savings scale directly with subscriber count.
* Full Acceleration assumes modern 100G-class nodes (~33k subscribers each at ~3 Mbps). On the existing 56 Gbps servers, Accelerated and Full both reach ~18.7k subscribers/node (network-card limited) — either way, roughly half the nodes of Standard.
Rack space & power — the saving that recurs every month
Fewer gateway nodes isn't only a one-time hardware saving — it permanently shrinks two recurring costs: the rack space you rent and the electricity you burn (the servers plus the cooling to carry their heat away). For the same 100,000 subscribers:
Rack space used (rack units "U" — lower is better)
Half the footprint frees rack you can rent out or grow into — without a bigger room.
| Mode | Nodes | Server power | + Cooling (facility) | 5-yr electricity * |
|---|---|---|---|---|
| Standard | ≈ 10 | ≈ 4.0 kW | ≈ 6.4 kW | ≈ $42,000 |
| Accelerated QoS | ≈ 5 | ≈ 2.0 kW | ≈ 3.2 kW | ≈ $21,000 (−50%) |
| Full Acceleration | ≈ 2–3 | ≈ 2.1 kW | ≈ 3.4 kW | ≈ $22,000 (−48%) |
* Illustrative: ~$0.15 per kWh, cooling overhead ×1.6 (typical data-centre). Accelerated nodes draw less because there are fewer of them; Full nodes are individually beefier but far fewer. Both roughly halve power and rack vs Standard — and these costs recur every month for the life of the deployment.
The bottom line
For the same number of subscribers, moving from Standard to acceleration can cut the gateway hardware you need — and the power, space and maintenance that come with it — by roughly half (Accelerated QoS) to as much as five-fold (Full Acceleration), while making the service faster for the customer.
Recommendation: Accelerated QoS is the safe, high-return default — about half the cost with every feature kept. Full Acceleration is the choice where scale, density and the lowest possible cost per subscriber matter most.