When AI models spill beyond a single GPU, the network connecting the processors becomes part of the computing engine. Nokia says tuning that network and how software uses it raised throughput in one two-node molecular-dynamics test by almost 34%, with the same GPUs in place.
Nokia Watch analysis
What happened
In a Nokia engineering post, Mohammad Zaman describes tests of high-performance computing and AI workloads across GPU clusters. Nokia’s testbed uses nodes with eight GPUs each, connected using RoCEv2 networking.
The post gives two results. For HPCG, scaling from one eight-GPU node to two nodes with 16 GPUs raised aggregate throughput from about 5.49 to 10.0 TFLOP/s, which Nokia calls 1.82× improvement and about 91% scaling efficiency. In a separate LAMMPS simulation of 2.048 million atoms on 16 NVIDIA H200 GPUs, throughput rose from 821 to 1,105 timesteps per second, almost 34%. Nokia says the compute resources stayed constant; the change was how MPI and UCX communications accessed the network interfaces.
Key findings
- More GPUs brought more HPCG throughput
Nokia measured 10.0 TFLOP/s across 16 GPUs, compared with about 5.49 TFLOP/s on eight. - Communication tuning improved the LAMMPS result
Nokia reports nearly 34% more timesteps per second without changing the GPUs. - AI and HPC share a networking headache
Nokia says both depend on fast, predictable communication as work is distributed across processors.
Why it matters
Buying more accelerators is only part of scaling AI or scientific computing. When a model is split across GPUs, or a simulation is divided among nodes, processors must exchange data and coordinate their work. Delays in that traffic can leave expensive hardware waiting around, which is an inelegant way to spend a data-centre budget.
These are Nokia’s own test results, not an independent comparison of network vendors or a guarantee that every workload will see the same gains. Still, the LAMMPS example makes a useful point: with the GPUs held constant, software and network configuration can materially change application throughput.
Our read
The interesting result is not simply that a faster network helps. It is that application-level performance depends on how computing software, GPU-to-network connections and the fabric work together. Nokia’s measurements make that case more concretely than another chart of peak bandwidth.
For teams running distributed AI or scientific workloads, the sensible takeaway is to benchmark the actual application and its communication patterns, not just compare headline network speeds. The network may be part of the compute stack, but that does not make every vendor’s best-case result your best-case result.
What to watch
- Whether Nokia publishes more detail on test configurations and repeatability.
- Whether the reported gains hold across other workloads, cluster sizes and network setups.
- How application-level performance compares across different fabrics and tuning choices.
Discussion spark: When scaling AI across GPUs, should teams prioritise faster accelerators or spend more effort tuning the network and software that connect them?
Sources and evidence
- Source update (9 November 2026, 21:55 UTC)
Independent WittyWires tracker for public updates about Nokia. Not affiliated with or endorsed by Nokia; this is not an official account.