Loading...
Technical Insight

Magazine Feature
This article was originally featured in the edition:
Issue 5 2026

Exorcising the DSP: A back-to-basics route to co-packaged optics

News

Every high-speed optical module hides a digital signal processor that quietly eats power and adds delay. Mark Holm argues the way forward is simpler modulation, many wavelengths and a single comb source.

BY Dr Mark Holm, VP of Technology, Solinide Photonics AB


Ask an engineer what made the last decade of optical interconnect possible, and you will hear about lasers, modulators and silicon photonics. Ask what actually carried the signal across a lossy backplane, through two connectors and down a dispersive fibre, and the honest answer is the digital signal processor. The DSP equalises the channel, corrects the errors and lets us run lane rates that the raw physics would never tolerate on its own. It has been the quiet hero of every pluggable module since 100G.

It has also become the hidden cost. In a modern pluggable, the DSP is the hottest thing in the box, and in an AI datacentre, where every watt spent on the network is a watt not spent on compute, that matters. It adds delay as well as heat: the signal has to be sampled, equalised, decoded, and corrected before it goes anywhere. We have come to accept this as the price of speed. I think that is a mistake, and I think the industry’s own roadmaps are beginning to say so.

Datarate: The treadmill that feeds the devil
The orthodox way to add bandwidth is to run each lane faster. Fifty gigabits became a hundred, a hundred became two hundred, and four hundred is on the slides. Each step costs more than the one before. The electrical channel gets harder, the equaliser gets deeper, the error correction gets stronger, and the copper reach gets shorter, which is why the optics have been forced off the front panel and onto the package in the first place. The DSP grows at every step, because it is the DSP that makes the step possible.

There is a second way to add bandwidth that costs none of this. Instead of running one wavelength faster, run more wavelengths.

Dense wavelength division multiplexing has done this in long-haul networks for thirty years, and it is now arriving at the chip. Hold the lane rate at something the physics is comfortable with and multiply the lanes in colour rather than in clock.

The pleasant side effect is that the simple formats come back into play: non-return-to-zero, a one and a zero, no equaliser worth the name, no forward error correction, and a clock forwarded on its own wavelength, so the receiver never has to recover one. A link like that is about as close to a wire as optics gets.


Figure 1: The devil in the box: DSP as the hot spot of a pluggable module. Inside a DSP-based pluggable module the signal processor is the hottest component in the box. It equalises the electrical channel, recovers the clock and corrects errors, and every one of those jobs costs power and adds delay before a single bit reaches the fibre.

Latency: The cost nobody puts on the datasheet
The watts are visible. The nanoseconds are not, and in an AI fabric they may matter more. Training a large model means thousands of accelerators exchanging partial results and waiting for the slowest one to answer. Every hop through a DSP and an error-correction block adds delay, and the delays compound across the fabric. The simplest NRZ link with a forwarded clock has the shortest electrical path from one chip to another that we know how to build. It is no coincidence that the people building the largest fabrics are the ones talking about dropping the DSP.

Shoreline: Bits per millimetre of edge
Once the optics moves onto the package a new limit appears. The electrical reach from the accelerator to the optical engine is now measured in millimetres, so the engine has to sit right at the edge of the die, and there is only so much edge. Engineers call it shoreline, or beachfront, and it is becoming the scarcest real estate in the datacentre. The question is no longer how fast one fibre can go. It is how many bits you can push through each millimetre of edge.

Many wavelengths per fibre is the obvious answer, and here the devil comes back in through the side door. Many wavelengths has always meant many lasers, and each laser brings its own driver, its own cooler, its own monitor and its own control loop. Sixteen wavelengths means sixteen of everything. Forty-eight means forty-eight. The electronics you threw out with the DSP creep back in, one laser at a time, and the shoreline you saved on the fibre side you spend again on the source side.

The source: The saving is in what you do not build
This is where optical frequency combs change the arithmetic. A comb takes one pump laser and one microresonator, a ring of silicon nitride less than a millimetre across, and produces a whole grid of wavelengths at once, evenly spaced and locked to one another. One pump, one temperature loop, one control circuit, and as many wavelengths as the ring will give you.

It is worth being precise about where the saving comes from, because it is easy to get wrong. A comb is not a more efficient laser. Line for line, a good semiconductor laser turns electrical power into light at least as well as a comb does. The saving is in everything that sits around the laser. A laser array carries a stack of control electronics for every wavelength; a comb carries one stack for all of them. That is why even a modest comb can hold its own against an array of lasers at the system level, despite being the less efficient light source. The electronics it does not need pay for the light it wastes.

A high-efficiency comb goes further. When the comb converts pump light into usable lines efficiently, with enough power to keep the links error-free, the removal of the per-line electronics stops being a trade and becomes a straight gain. In our own system modelling, a co-packaged engine fed by an efficient comb comes out at roughly half the energy per bit of the same engine fed by a laser array, and the whole of that difference is electronics that were never built. Against a DSP-based pluggable, the gap is an order of magnitude, which is the gap the industry is actually trying to close.


Figure 2: Two scaling axes: lane rate versus wavelength count.Two ways to add bandwidth. Running each lane faster (vertical axis) demands deeper equalisation, stronger error correction and shorter copper reach at every step. Adding wavelengths (horizontal axis) multiplies capacity while leaving the modulation format, and the electronics, unchanged.

Why the industry’s own roadmap points this way
NVIDIA has published a roadmap towards sub-picojoule-per-bit optical interconnect that is worth reading twice. Its first conclusion is that once a link goes DWDM, with simple modulation and a forwarded clock, the analogue circuits stop being the problem and the optical power becomes the largest single slice of the budget. Its second conclusion is a list of what to fix: better detectors, better rings, better processes, direct drive from the die. Every item on that list attacks the receiver, the modulators or the circuits. The source is not on it.

I think the source is the list. If optical power is the biggest slice and per-wavelength electronics is the biggest hidden multiplier, then a single efficient source that feeds every wavelength is the lever the roadmap is missing.

One objection deserves a direct answer. Running NRZ without error correction is not free: it demands real optical margin on every wavelength, because there is no DSP to rescue a weak channel. That might sound like an argument against the simple link. It is the opposite. It is precisely why the quality of the source decides whether the simple link is possible at all. A comb that delivers its lines at the right power, at the right spacing, in the right band, and holds them there as the package heats and cools, gives every wavelength the margin it needs. A poor comb, or a laser array that drifts, forces you to put the electronics back. The source the roadmap leaves off its list is the thing that makes the rest of the roadmap work.


Figure 3: Laser array with per-line control stacks versus a single microcomb. A laser array carries a driver, cooler, monitor and control loop for every wavelength; a microcomb carries one set for all of them. A silicon nitride microresonator, pumped by a single laser, generates an evenly spaced grid of wavelengths that can be modulated and received by simple ring resonators with no DSP in the path.

Back to basics
So here is the vision. NRZ, no FEC, lots of wavelengths, one comb. The modulator is a ring, the receiver is a ring and a photodiode, the clock rides on its own colour, and the DSP leaves the module. The link becomes as dumb and as fast as a wire, which is what the accelerator wanted all along.

The roadmap axis changes with it. Instead of counting lane rate, we count wavelengths: a few tens from a single source today, approaching a hundred within the decade, with the electronics staying the same size while the bandwidth multiplies. The external laser source, the form factor the industry is now standardising, becomes the natural home for the comb, feeding several engines from one pump.

Thirty years ago, we built optical links with a laser, a photodiode and not much else, and we were rather proud of how simple they were. We added the DSP because the channel got hard. The channel is about to get easy again, because it is a few millimetres of interposer and a short run of fibre. It is time to let the devil go.



Logo
x