SiPM vs PMT: Why Silicon Hasn’t Won Everywhere Yet

Ask ten particle physicists whether SiPM vs PMT is still a real debate, and most will say silicon won years ago. I think that answer is lazy. I spent years reading out photon signals on CERN beamlines. Silicon photomultipliers have indeed won almost everywhere new detectors get designed, yet the old photomultiplier tube still beats them on the metric nobody puts on a marketing slide: collecting area per euro.

Most “SiPM vs PMT” explainers read like a spec sheet. This one gives the numbers, says where I would still specify a vacuum tube in 2026, and explains why the economics that favour silicon in particle detectors also shape the optics in data centres.

Why physicists are asking SiPM vs PMT at all

A photomultiplier tube is a vacuum device: a photon knocks an electron out of a photocathode, and a chain of dynodes multiplies that single electron into a measurable pulse. It has been the workhorse of particle and nuclear physics since the 1930s, and it still holds the record for collecting area. A single 20-inch PMT of the kind used in the neutrino experiment JUNO covers more liquid scintillator than a wall of silicon sensors could for the same price.

A silicon photomultiplier is an array of thousands of tiny avalanche photodiodes on one chip, each a single-photon avalanche diode (SPAD) running in Geiger mode. Instead of a fragile glass tube under vacuum you get a solid-state sensor the size of a fingernail. It survives being dropped, works next to a magnet and runs on a few tens of volts: Hamamatsu’s S14161 has a typical breakdown voltage of 38 V and is operated 2.7 V above it (Hamamatsu datasheet).

The numbers behind the SiPM vs PMT debate

On photon detection efficiency silicon wins. The S14161 reaches about 50% at its 450 nm peak (Hamamatsu). JUNO’s 20-inch tubes, among the best large PMTs ever built, give an expected mean of 30.1% for the central detector, measured at 420 nm (JUNO mass testing, 2022). The wavelengths differ slightly, but the gap is large: a SiPM-based detector needs less light or exposure time to reach the same statistical precision.

The voltage story is even more lopsided. PMTs need supplies of one to two kilovolts to accelerate electrons down the dynode chain, which means bulky, power-hungry hardware and a real safety consideration in a detector with thousands of channels. SiPMs operate at a few tens of volts, typically between about 25 and 60 V. I have built readout boards for both. A kilovolt distribution system is a subsystem in itself, with its own cabling, shielding and failure modes; a SiPM array needs a PCB trace.

Magnetic fields are where PMTs struggle most, because the field bends the electrons on their long path from photocathode to first dynode. Large tubes are disturbed even by the Earth’s field of about half a gauss. JUNO surrounds its central detector with active compensation coils to keep the residual field below 10% of the geomagnetic value (JUNO coil study). SiPMs are solid-state and essentially insensitive to such fields, which is why new detectors that sit close to a magnet now default to silicon.

Where the photomultiplier tube still wins

Area is the PMT’s stronghold. A 20-inch tube (508 mm across) presents about 2,000 cm² of photocathode in projection, while a large 6 × 6 mm SiPM offers 0.36 cm², more than 5,000 times less (our arithmetic). Matching that area with silicon means tiling thousands of devices, and every tile brings its own dark counts, readout channel and cost. For a huge, sparse volume such as a kiloton-scale neutrino tank, a wall of glass tubes still beats a wall of silicon on cost per square metre.

Dark counts used to be silicon’s other weakness. Early SiPMs produced around 1 MHz per mm² of spurious pulses; modern devices are below 100 kHz per mm², and the rate roughly halves for every 8 °C of cooling. At room temperature that is still far more noise per square centimetre than a PMT, which matters when counting very faint light. Cold detectors flip the picture: at liquid-argon temperature SiPM dark counts become tiny, and the DarkSide-20k dark-matter experiment reads out its argon with silicon photomultipliers (DarkSide-20k design).

The data centre parallel

SiPMs and the photodetectors in optical transceivers share the same junction physics: a photon absorbed in a semiconductor creates an electron-hole pair that becomes current. They are used differently. Most short-reach data-centre receivers use PIN photodiodes without internal gain, avalanche photodiodes in linear mode appear in some longer-reach optics, and Geiger-mode SPADs, the SiPM’s building block, are used where single photons matter, as in lidar. The common thread is the transition happening in both worlds: bulky, high-voltage, single-purpose hardware giving way to low-voltage silicon that scales by adding identical units.

That is the manufacturing curve at work. Once a photon-counting technology moves onto a CMOS-compatible silicon process it inherits the semiconductor industry’s cost trajectory, as thin silicon tracking detectors did. Vacuum tubes never get that ride down the cost curve, however good the glassblowing. It is the deeper reason SiPMs keep winning new designs, even where PMTs still outperform them today.

It also explains why the SiPM roadmap looks different. PMT improvements arrived over decades: a better photocathode here, a faster dynode geometry there. SiPM detection efficiency, dark count rate and cell density have improved much faster, as the manufacturer comparisons from AP Technologies and the readout benchmarks from Berkeley Nucleonics describe. I expect a shrinking set of niches where a 90-year-old vacuum tube remains the economically correct answer, rather than a sudden extinction.

What comes next for photon detection

I doubt the next decade produces a clean winner; I expect hybrids. Large-volume neutrino and dark-matter experiments at room temperature will keep using PMTs where area per euro dominates, while compact, magnet-adjacent, cryogenic and timing-critical applications keep migrating to SiPMs. The interesting frontier is fully digital SPAD arrays that timestamp individual photons on chip instead of summing current. They are not new: Philips described a digital silicon photomultiplier in 2009 (Frach et al., IEEE NSS 2009). Pairing that digital readout with large areas at low cost is where I expect the real change by the early 2030s.

What I would watch next is whether SPAD-based receivers find a role in data-centre optics, as avalanche photodiodes did in longer-reach links. If they do, the traffic between particle-detector electronics and data-centre hardware will run in both directions. That is the pattern worth watching, more than the tired “silicon eats everything” headline.

@online{scienceshot-sipm-vs-pmt-verdict-2026,
  author       = {Major, Albert},
  title        = {{SiPM vs PMT: Why Silicon Hasn't Won Everywhere Yet}},
  organization = {ScienceShot},
  date         = {2026-07-12},
  year         = {2026},
  url          = {https://scienceshot.com/post/sipm-vs-pmt-verdict}
}