SC26 News Rodeo: Quantum, 800G Builds Out and Optics Reaches Memory
17 hour ago / Read about 26 minute
Source:Converge Digest
At SC26 in Chicago, quantum processors moved deeper into HPC job schedulers, researchers distributed entanglement across city fiber, and 800G Ethernet and photonic memory showed how classical systems are changing.

Quantum processors are starting to show up in the same job queue as CPUs and GPUs. At SC26 in Chicago, Pasqal announced support for Slurm 26.11’s new resource-management interface for its neutral-atom processors, while IQM outlined a way to schedule superconducting systems through the same workload manager. An HPC center can use the new interface to reserve a scarce quantum processor for the portion of a workflow that needs it and release the device for the next job. That gives researchers a practical way to test hybrid applications on the infrastructure they already use.

Across Chicago, a Northwestern University team reported distributing entangled photons over about 35 km of installed fiber in a demonstration coordinated across three sites. It was one of two notably different quantum networking stories at the show. On a route linking Bilbao, Amsterdam and Chicago, Ciena, Luxquanta and their research partners demonstrated how local quantum key distribution and post-quantum cryptography could work together to secure high-speed optical transport over an approximately 8,000 km path.

The classical side of the show had its own measures of progress. The November TOP500 left all ten leaders in place, even as Oak Ridge opened an AI system built with 800G-capable Ethernet switching. On the show floor, Penguin Solutions proposed pooling high-bandwidth memory across racks, and Q.ANT showed an AMD GPU handing off Fourier transforms to an optical processor. Together, the announcements show where HPC builders are changing the system around the processor, from its scheduler and network to the location of its memory.

SC26 at a glance
• Quantum scheduling · Pasqal and IQM detailed separate routes to making quantum processors schedulable through Slurm 26.11.
• Chicago fiber · Northwestern reported distributing entangled photons across about 35 km of installed fiber in a demonstration coordinated across three sites.
• TOP500 · LineShine held No. 1 at 2.198 exaflop/s on HPL. Only 27 systems were new to the list.
• Optical transport · Ciena and Internet2 carried 800 Gb/s coherent pluggables across a 2,104 km production-network loop. A separate QuLore trial demonstrated hybrid quantum-safe security on an approximately 8,000 km route.
• Photonic systems · Penguin targeted 57.6 Tb/s aggregate bandwidth for pooled memory. Q.ANT demonstrated optical Fourier transforms in a hybrid language-model workflow.

Quantum processors enter the job queue

Slurm’s Hierarchical Resources interface, known as HRES, lets a computing center make a quantum processing unit available as a cluster-level resource. Pasqal is using the interface with the open-source Quantum Resource Management Interface, or QRMI, which it already supported. Slurm can check availability, allocate access and release a device when a job finishes. Pasqal says it verified the integration in a simulated on-premises environment and on Pasqal Cloud. At Italy’s CINECA, work is underway to connect a Pasqal processor to the Leonardo supercomputer through Slurm and QRMI.

“Quantum computers will not be adopted as islands. They will be adopted when they become one more resource in the infrastructure supercomputing centers already run,” said Loïc Henriet, CTO of Pasqal.

IQM is taking a parallel route through the Quantum Device Management Interface, or QDMI, using its partner MQSC’s orchestrator. It also announced an upgrade path for Halocene from 150 physical qubits to more than 1,000 physical qubits, with integration services covering existing schedulers and GPU systems. Finland’s LUMI AI Factory has selected Halocene for LUMI-IQ, which is planned to connect to the upcoming LUMI-AI supercomputer. The roadmap spans several hardware generations, while the scheduler work addresses how researchers can access the machines as they arrive.

IBM offered an example of what could run on this kind of infrastructure. In a project with Cleveland Clinic and RIKEN, classical systems divided protein–ligand complexes of up to 12,635 atoms into smaller computational problems. IBM Heron processors were used alongside classical HPC on selected harder subproblems. The full-complex size describes the hybrid workflow. IBM and NVIDIA also announced plans for a modular platform with Fermilab to investigate high-energy and nuclear physics workloads.

Two different tests for quantum networking

Northwestern’s SC26 demonstration coordinated equipment in Evanston, at the StarLight exchange and at McCormick Place. One photon from each entangled pair stayed in the Evanston lab while its partner traveled to StarLight. The team used automated correction to compensate for polarization drift in installed fiber. The experiment builds on peer-reviewed work in which the researchers distributed entanglement across 24.4 km of deployed fiber carrying high-capacity classical signals. Northwestern identified a future attempt to extend entanglement distribution to Argonne.

The separate QuLore demonstration used quantum key distribution on local links and post-quantum cryptography across the wider Bilbao–Amsterdam–Chicago topology. The organizers reported optical encryption at 1.6 Tb/s in Chicago and 800 Gb/s on the Bizkaia link. Those rates describe local encrypted optical channels. QuLore, developed by the University of the Basque Country, coordinated how the different mechanisms were applied. The trial is a case study in deploying quantum-safe encryption across different links and administrative domains.

Chicago also put its regional infrastructure on display. At the media briefing, Chicago Quantum Exchange founding director David Awschalom called this the “prove it” phase for quantum technology. He pointed to two distinct federal tests. The Department of Energy’s Quantum Genesis Q competition calls for systems with at least 100 logical qubits and hundreds of millions of fault-tolerant operations. DARPA’s Quantum Benchmarking Initiative asks whether any approach can deliver computational value greater than its cost by 2033. Logical qubits encode information across physical qubits so a system can detect and correct errors during computation. They give evaluators a different measure from the physical-qubit counts on hardware roadmaps.

The TOP500 holds steady as fabrics diversify

LineShine remained No. 1 on the November TOP500 at 2.198 exaflop/s on High Performance Linpack, followed by El Capitan at 1.809 and Frontier at 1.353. Only 27 of 500 systems were new to the list, a record low, while aggregate HPL performance rose 4.4% from June to 19.57 exaflop/s. The Green500’s top ten also held steady, with KAIROS first at 73.28 gigaflop/s per watt. El Capitan led the separate HPL-MxP mixed-precision ranking at 16.7 exaflop/s. Each result measures a different aspect of the machines.

Oak Ridge’s Lux debuted at No. 14, reaching 220.79 petaflop/s on HPL with AMD Instinct MI355X GPUs. Its HPE Juniper QFX switches support 800 Gigabit Ethernet. Across the show, NVIDIA presented a 192-CPU Vera rack linked by Quantum-X800 InfiniBand, and Cornelis showed its 800G CN6000 portfolio with Omni-Path and RoCEv2 offerings. ETH Zurich’s Torsten Hoefler argued at the briefing that Ethernet has room to converge with HPC fabrics as transport and congestion control improve. His example was Microsoft’s Multipath Reliable Connection work, which used packet spraying and fast recovery to keep a large training job running through a switch failure. The systems on display show how much variation remains in fabric choices.

The show’s own network provided the wide-area counterpart. Internet2 said SCinet planned 19.6 Tb/s of wide-area capacity, including 7.6 Tb/s from Internet2. That contribution included 2.8 Tb/s from Seattle directly to the StarLight booth, assembled from two 800 Gb/s and three 400 Gb/s circuits. In a separate test, Ciena and Internet2 ran 800 Gb/s coherent pluggables over a 2,104 km production-fiber loop using 135 GBaud signaling and EDFA-only amplification.

Photonics finds two jobs in the rack

Penguin Solutions’ MemoryAI Photonic Appliance uses Marvell Photonic Fabric to give CPU and GPU servers access to a shared pool of up to 16 TB of DDR5 and 576 GB of HBM3E. Penguin targets 57.6 Tb/s of aggregate bandwidth and plans select-customer sampling in the second half of 2027. The immediate question for inference operators is how much frequently reused data, including a model’s key–value cache, can live in a shared pool while meeting application latency requirements.

Q.ANT used light for a different purpose. Its NPU Gen 2+ performed discrete Fourier transforms as part of a language-model demonstration with an AMD GPU handling other work. Q.ANT plans commercial systems combining its Gen 2 and Gen 2+ processors in the second half of 2027. The next comparison to watch is the end-to-end cost of the hybrid workflow, including the work of handing data between the GPU and the optical processor.

Analysis

The strongest thread through SC26 was the work of fitting new hardware into systems researchers already operate. For quantum processors, the scheduler makes it possible to measure whether a hybrid job gains enough from an accelerator to justify queue time, data transfers and repeated measurements. For shared memory, the comparison is application throughput and latency after the workload crosses the fabric. For classical clusters, the HPL standings offer a stable reference point, while applications expose differences in precision, memory traffic, power and resilience. The next round of results will be most instructive when it measures the complete workflow.

Company profiles · Pasqal and IQM
PasqalPalaiseau, France · Founded in 2019 · CEO Wasiq Bokhari · Nasdaq: PSQL
TechnologyNeutral-atom quantum processors with on-premises and cloud access. Pasqal works with Slurm and QRMI for hybrid workflows.
IQM Quantum ComputersEspoo, Finland · Founded in 2018 · Co-founder and CEO Jan Goetz · Nasdaq: IQMX
TechnologySuperconducting quantum computers for on-premises HPC integration and cloud access. Platforms include Halocene and IQM Resonance.