3.2 Pb/s in real time: what China's 24-core fiber record does and doesn't show

A Chinese consortium reports that it has moved 3.2 petabits per second through a single 24-core fiber. The test was real-time and bidirectional, and the team calls it a world record. The work was led by the National Key Laboratory of Optical Communication Technology and Networks at China Information and Communication Technologies Group (CICT), with Pengcheng Laboratory and fiber maker FiberHome Fujikura. The link was 10.3 km of multicore fiber carrying traffic in both directions at once.

The headline rate is less informative than three qualifiers: real-time, 10.3 km and bidirectional. Together they say what the demonstration proves and what it leaves open.

01What was built

Multicore fiber puts several light-guiding cores inside one cladding. A Global Times interview with a CICT researcher, from the earlier 2.5 Pb/s result, compared conventional fiber to a single lane and 24-core fiber to 24 lanes in one strand. The team stacks this spatial dimension on top of wavelength multiplexing. It combined the S, C and L bands with space-division multiplexing, which yields thousands of parallel channels.

The team gave a few line-rate figures. Its own optical modules ran 400 Gb/s per wavelength in the S band and 1.2 Tb/s per wavelength in the C and L bands. Reports say the modulation used high-order probabilistic shaping. Probabilistic shaping adjusts how often each constellation point is used, so the link can run close to the capacity of each channel's signal-to-noise ratio. The S band gets a lower rate, which fits its generally harder amplification and loss conditions. I could not find the full channel plan, amplifier types or per-core power budgets in the public material.

The team's own comparison is with ordinary single-core G.652 fiber. It says 3.2 Pb/s is 32 times the capacity of such systems, which typically reach about 100 Tb/s using the C and L bands. The arithmetic holds: 32 × 100 Tb/s is 3.2 Pb/s. Dividing 3.2 Pb/s by 24 cores gives roughly 133 Tb/s per core. That is my calculation, and it presumably counts both directions. It is a little above the C+L-only figure for a single-core system, which is consistent with adding S-band spectrum.

02Why "real time" is the important word

Many record-setting transmission experiments capture waveforms and process them offline. Researchers record the received signal and then apply heavy digital signal processing on lab equipment afterwards. Offline processing lets a team use algorithms that no commercial module could run within its power and latency limits, so such results show what the physics permits. This result instead used commercially compliant hardware architectures running in real time.

Real-time operation means the following happen inside module-class silicon:

• Coherent detection and equalization of chromatic dispersion and polarization effects
• Carrier recovery
• Forward error correction decoding
• Probabilistic-shaping demapping

All of this runs continuously, within a thermal budget a pluggable or line-card module can tolerate. The ceiling is usually lower than offline numbers, and that gap is visible here. Lab results on similar 24-core fiber have reached up to 5.29 Pb/s, according to Tom's Hardware, which gives no underlying paper. Further afield, offline experiments in other fiber types are far higher. NICT and partners reported 22.9 Pb/s from a 38-core, 3-mode fiber in 2023. That used a MIMO receiver, a different architecture from a plain multi-lane link.

03The hard parts: crosstalk and band heterogeneity

The reports name two main obstacles. One is inter-core crosstalk. The other is meeting the differing requirements of the S, C and L bands.

Crosstalk is the leakage of light between neighboring cores, and it grows with core density and with wavelength, because longer wavelengths are less tightly confined. A 24-core layout is a compromise between packing cores in and keeping them isolated. If leakage is low enough, each core can be treated as an independent channel and decoded with standard coherent DSP. That avoids the complexity of joint multi-channel equalization, which is hard to do in real time. The Asia-Pacific Fiber Optic Cable Industry Association has just published what it calls the first international standard for weakly coupled single-mode multi-core fiber. I have not seen the standard, but the term "weakly coupled" points to this independent-channel design. Reports say the team used modulation techniques to limit inter-core interference, but they do not say how.

The three bands are the second problem. Fiber attenuation, nonlinear behavior and amplifier gain differ across them. S-band amplification is less mature than the erbium-based C and L bands. Transceivers have to be tuned to each band, which is consistent with the lower 400 Gb/s per-wavelength S-band rate. Multicore fiber adds a further burden: amplifiers and test gear must serve 24 cores at once.

04Distance, scrutiny and where it could land

The record link is short. At 10.3 km, the distance says a lot about where the technology would land first. That range suits campus and data-center interconnects, where fiber is expensive to pull, conduit space is tight and aggregate bandwidth matters more than reach. A backbone or subsea route is a much harder case. Long haul requires inline amplification for every core, and subsea cables are limited by power delivery. Submarine cables are practically limited to around 24 fiber pairs because the amplifiers are powered from shore. Multicore fiber is one route around that limit, but this experiment says nothing about amplified spans. For comparison, NICT sent 1.02 Pb/s across 1,808 km of 19-core fiber last year, according to Hardware Busters. That is a far longer reach at a third of the rate.

Several caveats apply:

• Peer review is pending. I found no primary technical paper or release with the full experimental details. The most direct sources are Chinese-language coverage, secondary reports and an earlier conference abstract. Reports say the results were invited as a paper to the Journal of Lightwave Technology, which is where independent scrutiny will happen.
• The "record" is narrowly scoped. It applies to real-time transmission over 24-core fiber, not to fiber capacity in general.
• The timeline is uncertain. Reports give the earlier mark as 2.5 Pb/s earlier this year. CICT announced real-time 2.5 Pb/s bidirectional transmission over 24-core fiber in March. The jump is about 28%, since 3.2 divided by 2.5 is 1.28. Some outlets describe it as a gain "within a year", and the actual interval was likely a few months.
• The consumer comparisons are only arithmetic. The 20,000 4K movies per second and 3.2 million gigabit households claims just restate the rate. They say nothing about deployability.
• Deployment cost is unaddressed. The reports give no figures for connectors, splicing, amplifier cost or cable manufacturing yield. They also give no per-bit power figures.

The result is best read as evidence that multicore SDM and S+C+L transmission can run on module-class real-time hardware at petabit scale over short distances. It does not show that these links are economical against simply lighting more single-core fibers. The next evidence to look for is the Lightwave Technology paper, a field trial in installed ducts and amplified, longer spans. Field-deployed fiber faces aging, splice losses, vibration, temperature swings and optical power limits that a lab span avoids.