NVIDIA's SC26 Lineup: Rubin HPC GPU, 192-CPU Vera Racks and Quantum Decoding
NVIDIA's SC26 slate includes a Rubin HPC GPU due in 2028, 192-CPU Vera racks in early 2027, CUDA-X updates and quantum decoding claims.

The headline item is Rubin HPC, a GPU aimed at simulation and data processing rather than AI training. NVIDIA is pitching it as the scientific-computing member of the Vera Rubin platform, with more than 50,000 CUDA cores on one package and availability expected in 2028. The rest of the SC26 announcements point the same way: CPU racks for HPC sites arriving in early 2027, a new open model package for science, updated CUDA-X libraries, quantum tooling, and a deployment of co-packaged optics. If you run clusters, the timing matters more than the spec sheet.

Rubin HPC: Compute Plus Data Movement
NVIDIA describes Rubin HPC as a GPU for simulation, data processing, and analytics in science, engineering, and finance. The design goal is to move data as fast as it computes. Beyond the core count, which NVIDIA calls a record, the company points to sparse data acceleration that keeps those cores supplied with data, and GPU-to-GPU links over seventh-generation NVLink. The target workloads are sparse matrix operations, stencil computations, sorting, and reductions, which are the patterns underneath climate models, engineering simulation, genomic analysis, and risk calculations.

On stage, NVIDIA said Rubin HPC triples the performance of simulation and analytics and supports every precision from FP64 to FP4. Here's the thing: the company didn't name the baseline for that 3x figure, and it hasn't published memory capacity, FP64 throughput, power draw, or system configurations. Until those numbers appear, any sizing exercise is guesswork. Ian Buck, vice president of hyperscale and high-performance computing at NVIDIA, tied the part to agentic research, saying "faster simulation means faster discovery" for agents working on science and engineering problems. Intel is pitching its FP64-capable GPU D270 at the same simulation workloads at the show, so expect a side-by-side comparison once both parts have published specifications.
Vera CPU Racks for HPC
The HPC configuration of Vera is a liquid-cooled NVIDIA MGX rack holding 192 Vera CPUs, connected by Quantum-X800 InfiniBand. NVIDIA quotes about 300TB of memory, 1PB of storage, and more than 200TB/s of aggregate peak memory bandwidth for the rack. Each 1U tray carries eight CPUs with NVMe storage and management, and the rack reuses the chassis, power distribution, and cooling of NVL72 systems. Each CPU has 88 Olympus cores, 176 threads, and SOCAMM2 LPDDR5X memory rated at up to 1.2TB/s.

This isn't the same rack as the 256-CPU Vera racks Dell and Supermicro announced for agent sandboxes at GTC Berlin. The HPC version adds InfiniBand and SHARP in-network computing for tightly coupled simulation. NVIDIA says Dell Technologies, HPE, Lenovo, and Supermicro will offer these HPC racks starting in early 2027. DKRZ, the German Climate Computing Center, has selected Vera CPUs for its Bull-built HLRE-5 supercomputer, which also includes GB200 systems and Quantum-2 InfiniBand and is expected in early 2028.
NVIDIA claims twice the energy efficiency of x86 CPUs for Vera. In its own tests, Vera delivered 1.9x the performance and 2.4x the performance per watt of x86 systems on the OpenFOAM motorBike Large case. It projects 1.6x the performance and 2x the efficiency on the Weather Research and Forecasting model. These are vendor figures we haven't run or audited, and NVIDIA didn't name the x86 processors it compared against. For context, our own Vera CPU review found a two-socket Vera system finished a large OpenFOAM mesh 1.44x faster than a 384-core EPYC 9965 system and 1.20x faster than a 240-core Xeon 6978P system. Those are different test conditions, but they point in the same direction.
On software, NVIDIA HPC SDK 26.11 is available now, with compilers and libraries tuned for Vera. A next-generation Fortran compiler, built with HPC centers and national labs, is in public beta. NVIDIA says it runs existing Fortran programs faster on its platforms without a code rewrite. That's the detail most HPC teams with legacy codebases will want to test first.
Nemotron for Science and CUDA-X
Nemotron for Science bundles the Nemotron 3.5 Super VL open model with scientific evaluations, post-training recipes, and LitmusBench v0.2, an open dataset of nearly 50,000 chemistry reasoning questions. The model and dataset are available now on Hugging Face. NVIDIA says the model has been optimized to run locally on a 128GB DGX Spark. Argonne National Laboratory is integrating it into its StructBioReasoner multi-agent system for protein-based therapeutics, and RIKEN has made it available on its RIKYU AI supercomputing platform.
Customer results are mostly vendor-reported. Lila Sciences nearly doubled quantum-dot synthesis accuracy to 61%, which NVIDIA says beats GPT-6 Astra's 54% at 10x to 20x lower inference cost. Terray Therapeutics recovered about 76% of annotated reactions on the RxnScribe test for reading chemical reaction diagrams. A KTH Royal Institute of Technology agent repaired 81% of broken GROMACS molecular dynamics setups in one overnight run on a single DGX Spark.
The CUDA-X additions are where the speedup claims get large. The cuPhoton toolkit correlates Event Horizon Telescope observations in eight minutes, where the analysis previously took more than a month. cuPotential speeds existing molecular simulation software by up to 3x. Daviator, a set of batched GPU kernels for the Vienna Ab initio Simulation Package, runs quantum-mechanical calculations up to 22x faster at up to 75% lower cost. NVIDIA didn't state baselines for the cuPotential and Daviator figures, so treat them as directional. cuSpa and a GPU-accelerated HMMER library are now open source. cuSpa delivered a 900x speedup over CPUs on cell polygon calculations at Memorial Sloan Kettering Cancer Center, and the HMMER library ran benchmark searches up to 84x faster than standard HMMER at Basecamp Research. The rapids-singlecell library, co-developed with scverse, now ships in one container built for agents, and 10x Genomics has integrated it into its Sentira platform.
Quantum Codesign and Decoding
NVIDIA and IBM are colocating DGX B200 systems with IBM quantum computers, including systems running the IBM Quantum Nighthawk processor. The aim is a modular quantum codesign platform developed with Fermilab, where researchers can test error-correction codes, decoding algorithms, and quantum-classical applications for high-energy and nuclear physics. IBM has a similar arrangement with AMD and Oak Ridge National Laboratory, pairing Nighthawk with an Instinct MI355X cluster. NVIDIA gave no system counts, location, or timeline.
On the software side, NVIDIA released TCBP-JOBS, a decoding algorithm in the CUDA-Q QEC 0.9.0 library. The company says it decodes up to 13,000x faster than leading existing algorithms and cuts the computation behind new quantum computer designs from 12 weeks to 45 minutes. CUDA-Q also gains a decoder server, an expanded set of fault-tolerant primitives, a CUDA-Q Algorithms library of reusable components, and CUDA-Q Pulse for shaping the control pulses that drive qubits.
The scheduling angle may matter more to operators. NVIDIA says quantum processor deployments at supercomputing centers have grown nearly sevenfold in five years, with nearly 30 centers now running more than 50 processors. Slurm 26.11, due later in November, adds a quantum interface that lets a QPU builder describe hardware once, in a single file, and reuse that integration at every center. It supports the QRMI and QDMI quantum-HPC interfaces, with IBM contributing through its Quantum Resource Management Interface, so centers can schedule quantum processors alongside CPUs and GPUs. Alice & Bob, IBM, IQM, Pasqal, Quantinuum, and MQSC are adopting it. NVIDIA acquired Slurm developer SchedMD in late 2025.
NVIDIA also cited quantum-classical chemistry work by IBM and RIKEN, a finalist for the 2026 ACM Gordon Bell Prize. The team calculated binding energies for protein-ligand complexes of up to 12,635 atoms, more than 40x the previous 303-atom quantum-classical benchmark. The classical side ran on RIKEN's ROQUO supercomputer, built on NVIDIA GB200 systems and Quantum-X800 InfiniBand. Ecosystem work includes an FPGA-based quantum controller from Altera and Oxford Quantum Circuits that links qubit control and readout to GPUs through NVQLink. QunaSys used cuQuantum and 1,024 Hopper-generation GPUs on ABCI-Q to simulate more than 300 million quantum operations, reaching chemical accuracy for a four-hydrogen-atom model. NVIDIA researchers used the JUPITER supercomputer to simulate 104 qubits under realistic noise.
Lambda Runs Co-Packaged Optics at Scale
Lambda is using NVIDIA Photonics co-packaged optics (CPO) switches to connect GB300 NVL72 racks in a deployment of more than 10,000 GPUs. Each switch has 144 ports at 800Gb/s and 18 external light sources that technicians can replace in the field. That matters for operations: with pluggable optics, a failed laser means swapping the whole module. NVIDIA cites testing that showed network latency about 4% lower at rack scale and up to about 10% lower at pod scale, with throughput matching pluggable optics.

In the Lambda deployment, median optics temperature was 36°C, against 51°C for conventional pluggable optics. NVIDIA rates its photonics switches at up to 5x better power efficiency, up to 15x higher reliability, and 4x fewer lasers than pluggable designs. Those are vendor ratings, and a single site won't settle them. NVIDIA also marked 20 years of CUDA, which it says now has more than 11 million developers and more than 1,000 CUDA-X libraries.
My read: the near-term actions for most teams are the 2027 Vera HPC rack planning, the HPC SDK 26.11 and Fortran beta testing, and watching for Slurm 26.11 if you run quantum workloads alongside GPUs. Rubin HPC's real numbers are still missing, so hold off on capacity commitments until NVIDIA publishes its baselines.