Every step up in port speed has added a digital signal processor inside the optical module, and that processor now consumes a noticeable fraction of the power budget of a switch. The proposed fix is to remove it and let the switch silicon drive the optical link directly. That saves watts and nanoseconds, and it also deletes the component that was cleaning up the channel.
A retimed module receives a degraded electrical signal from the switch, equalises it, recovers the clock, re-encodes and retransmits. That digital front end is why a conventional 800G module can tolerate a lossy PCB trace and a marginal connector. It costs roughly 14 watts per port in current designs, contributes real latency, and adds a firmware and interoperability surface that operators have to qualify.
Removing the retimer does not remove the signal-integrity problem; it moves the cost from a chip in the module to the board, the connector and the qualification process.
Part One: What is actually being compared
The two options are often described as competing implementations of the same thing, which is misleading. linear-drive pluggable optics and DSP-retimed pluggable optics make different promises about where the variability sits: one puts it in the capital plan, the other in the operating tail. A comparison that only measures the mean therefore misses the entire reason a team would choose one over the other in AI cluster optical links.
There is a quiet accounting asymmetry here. linear-drive pluggable optics moves spending into capital and spreads it over the life of the installation; DSP-retimed pluggable optics keeps spending in operations where it is easy to approve and hard to cap. Finance teams prefer the second, engineering teams prefer the first, and the argument is usually settled by the budgeting process rather than by measurement in AI cluster optical links.
Part Two: The arithmetic that decides it
A linear pluggable removes that chip and drives the laser or modulator with an analogue signal straight from the switch ASIC. Power falls to around 8 watts per port and the deterministic latency drops by a few nanoseconds. But the link budget that the DSP used to repair is now a hard constraint: trace length, connector loss, skew and temperature drift all have to close without a digital correction stage.
Deployment also changes the measurement. After linear-drive pluggable optics goes in, the metric people optimise shifts from throughput to predictability, and after DSP-retimed pluggable optics it shifts from predictability to unit price. That shift is why two teams can look at the same production data and reach opposite conclusions about which option won.
Part Three: Where the simple model breaks
This changes who owns the failure. With a DSP inside the module, an intermittent link is usually a module problem that can be fixed by swapping a part. With a linear link, the same symptom may come from the host board, the cage, the cable or the ASIC equalisation settings, and no single vendor owns the whole channel. Interoperability testing moves from a pairwise module check to a full-channel qualification.
Going from 14 to 8 watts per 800G port across a 3-hop fabric with hundreds of links changes the thermal design of the whole rack.
Part Four: The cost that arrives later
There is a second-order effect on network design. Lower power and lower latency make it economical to run more optical hops inside a rack, which is attractive for scale-up fabrics where every hop counts. But a shorter reachable channel also pushes designers toward co-packaged optics or shorter, cleaner board routes, which reduces the flexibility of the physical layout. The saving in watts is partly spent as a constraint on topology.
Adoption will not be a single flip. Early movers tend to be the cases with the most extreme value of the deciding variable, then the middle of the market follows once tooling and reference designs exist, and the long tail stays with DSP-retimed pluggable optics until the cost of switching falls below the cost of an exception. A reasonable reading of the next few years is coexistence with a slowly moving boundary rather than a replacement event.
Part Five: What this changes outside the engineering team
The economics therefore depend on how many ports are in one failure domain. In a 3-hop scale-up fabric with hundreds of links per node, saving six watts per port is a meaningful facilities and thermal win. In a small campus switch with a few uplinks, the same saving is not worth the qualification risk. Volume matters more than the per-port number in a datasheet.
Noise-cancelling headphones are the everyday version. The DSP version cleans up a bad acoustic environment inside the device; the linear version assumes the environment is quiet and is smaller, lighter and cheaper when it is. Take it onto a noisy train and the assumption is the whole story.
A market sizing sanity check
Optical modules for AI and cloud switching are roughly a 12 to 18 billion dollar annual market, growing faster than the rest of the networking stack. If linear-drive links take 20 to 35 percent of new high-radix ports by 2030, the displaced DSP content is a 2 to 4 billion dollar shift, while the additional spend on channel qualification, cleaner boards and host silicon equalisation moves a comparable amount of value to switch vendors.
All figures here are an author's synthesis of published ranges and a stated model, not a measurement of a shipped system. The point of the arithmetic is to locate the crossover and expose which assumption drives it; a vendor quote will differ, and the direction of that difference is more informative than the final number.
The comparison that matters
DSP-retimed optics buys a forgiving channel and a clean vendor boundary at the price of watts and latency. Linear-drive optics buys power and determinism at the price of a shared, tightly specified channel. Decide at the fabric level, where port count and hop count are known, not at the module level.
A change in regulation or in the supply of a single component could invalidate the ranking overnight. The ranking is a snapshot of prices and constraints in AI cluster optical links, not a property of the technologies, and anyone reusing it in a different year should re-run the threshold rather than reuse the conclusion.
The vendor boundary is a real engineering feature, and linear links dissolve it.
SOL: HCqPmPJM3NVtS6YrmVaTZYnuP8XZ2PFgQZKT94VK1LkS
Creating this takes real effort — thank you for the support.
Top comments (0)