2026-08-31
Nokia Bell Labs runs distributed LLM inference over 2-core and 4-core MCF, mapping TP and PP onto spatial channels; latency numbers sit in the unretrieved two-page body.
When an LLM is split across GPUs with tensor parallelism (TP) and pipeline parallelism (PP), activations have to move between devices. Inside a rack, NVLink can still hide that traffic behind compute. Across racks or buildings, fiber propagation delay eats the overlap window. Optical papers already show LLM training field trials over 120 km OTN, and separately sell 4-core and 16-core multi-core fiber (MCF) as a density play for data halls. What was missing is treating "how the model is sliced" and "how many cores sit in the cladding" as one mapping: parallelism strategy onto spatial channels, exercised with a real inference load.
Nokia Bell Labs and Sumitomo Electric filed a two-page SUM 2026 talk on that mapping. It is a photonics experimental note, not a serving-systems paper. The underlying tension is real. Inference cares about tail latency more than training does. PP point-to-point traffic and TP AllReduce do not tolerate delay the same way, and a single SMF pair gives the scheduler no spatial isolation.
The experimental topology is in the paper figure, not only in the abstract. Four nodes, each with a 100GbE NIC and a 100G DR pluggable. Three media run in parallel: a conventional single-mode fiber (SMF) pair drawn with a spool, a 2-core MCF, and a 4-core MCF.
The same figure splits two jobs:
The design choice is concrete. One 4-core fiber spatially multiplexes two independent distributed-inference jobs, so LLM-side parallel axes line up with fiber cores instead of dumping every GPU flow onto one SMF pair. Sumitomo makes the fiber. Citations include their 2019 field-deployed MCF testbed, a 4-core O-band MCF (100 μm cladding, 160 μm coating, about 9.8× core density versus single-core), and a 16-core MCF aimed at 1.6 Tbps data-center links. The software reference is Megatron-style model parallelism.
Which model, what parameter count, which collective library, how many kilometers on the spool, and how they measured crosstalk are in the two-page PDF behind the IEEE paywall. 100G DR is a short-reach spec, usually hundreds of meters. The spool in the figure implies they stretched the link. The actual span is in the unretrieved body.
The official abstract locks two claims: they study optical-interconnect latency under different parallelism strategies, and they experimentally demonstrate distributed LLM inference over MCF.
| What can be confirmed | Source | Number |
| Venue and length | CrossRef | SUM 2026, pages 1–2 |
| Nodes and optics | Paper figure | 4 nodes, 100GbE + 100G DR |
| MCF | Paper figure | 2-core and 4-core; 4-core carries two jobs |
| Parallel mapping | Paper figure | intra-node TP, inter-node PP |
| Throughput / latency / speedup | Body not retrieved | not in the abstract |
The paper does not give a comparison table versus InfiniBand, versus SMF-only, or versus simulation in any source retrieved here. Those numbers sit in the unretrieved text.
The practical question is whether campus-scale inference can put PP point-to-point traffic and TP collectives on different cores, trading spatial parallelism for cable density and job isolation. If that mapping holds, MCF stops being "four fibers in one jacket" and becomes a spatial channel the serving stack can see.
This is an incremental physical-layer demo. Four nodes and at most two GPUs per node are orders of magnitude below a production inference cluster. It shows the link can run, not that operators should deploy this way. Training already has 120 km, thousand-GPU field trials. Inference is starting to run on MCF instead of optical simulation only. Serving engineers will not pick this up next week. People looking at scale-across optics and MCF plant will recognize the topology.
The full text was not retrieved. Latency curves, model size, fiber length, BER, and baselines are unknown, so circulating distance figures from social posts do not belong in the results. A two-page SUM paper cannot fully cover a parallelism sweep, inter-core crosstalk, and end-to-end token latency at once. 100G DR was not designed for multi-kilometer campus spans; without extra optical compensation or a different pluggable in the body, distance claims should stay conservative. The setup is small, and there is no visible fair comparison against in-NVLink TP or Ethernet PP. Author-stated limits are in the unretrieved body.