Nvidia opened the SC26 supercomputing conference in Chicago on November 16 by announcing Rubin HPC, a GPU aimed at scientific simulation rather than AI training, which it says will pack more than 50,000 CUDA cores and arrive in 2028. It also announced a liquid-cooled rack of 192 Vera CPUs for early 2027. The same day, the 68th TOP500 list of the world's fastest supercomputers came out with no change at all in its top 10. China's LineShine, a system with no GPUs at all, is still No. 1.
Why it matters: Nvidia, whose recent GPUs have been sold mainly as AI hardware, is now pitching a separate GPU and a CPU rack built for simulation work. Meanwhile the fastest machine on the list shows that CPUs alone can still win the classic benchmark.
Rubin HPC: one big number, many gaps
On its blog, Nvidia describes Rubin HPC as part of the Vera Rubin platform. Its target is the "sparse" work common in climate, engineering, genomics and financial-risk codes: sparse matrix maths, stencil calculations, sorting and reductions. Nvidia's claims:
- More than 50,000 CUDA cores on one GPU, which Nvidia calls a record
- A sparse data acceleration feature that it says keeps those cores supplied with data
- Seventh-generation NVLink for links between GPUs
- Expected in 2028
What Nvidia didn't say matters as much. Neither the blog post nor the independent write-ups from StorageReview and Unite.AI give FP64 (double-precision) throughput, memory capacity or bandwidth, power, or any comparison with current Rubin parts or AMD's accelerators. Many physics and climate codes depend on FP64, so the core count says little about real performance. Unite.AI also notes that Nvidia doesn't say whether the sparse-acceleration gains were measured or are projections.
Vera CPU racks: early 2027, with named buyers
The Vera announcement comes with more detail, according to Nvidia's press release:
| Item | Nvidia's figure |
|---|---|
| CPUs per rack | 192 Vera, 8 per 1U tray |
| Cores per CPU | 88 Olympus cores, 176 threads |
| Memory bandwidth | Up to 1.2 TB/s per CPU (SOCAMM2 LPDDR5X); >200 TB/s per rack |
| Rack memory / storage | ~300 TB / 1 PB |
| Network | Quantum-X800 InfiniBand |
| Availability | Dell, HPE, Lenovo, Supermicro, starting early 2027 |
Nvidia claims 1.9 times the performance and 2.4 times the performance per watt of x86 systems on the OpenFOAM motorBike Large airflow test. For the WRF weather model it gives projected gains of 1.6 times the performance and twice the energy efficiency. The release doesn't name the x86 chips it compared against. StorageReview's own earlier Vera testing was more modest: a two-socket Vera system ran a large OpenFOAM mesh 1.44 times faster than a 384-core AMD EPYC 9965 system and 1.20 times faster than a 240-core Xeon 6978P system.
There is one concrete customer. Germany's climate computing centre, DKRZ, chose Vera CPUs for its Bull-built HLRE-5 system, alongside GB200 nodes. Nvidia says HLRE-5 should be available in early 2028. The company also shipped HPC SDK 26.11 and opened a public beta of a new Fortran compiler. Nvidia says that compiler can speed up existing Fortran programs without rewriting them.
TOP500: a frozen top 10
According to the TOP500 list page, the November 2026 top 10 is identical to June's. TOP500's own headline for the release says the rate at which new systems enter the list has fallen to a historic low. The announcement text with the turnover numbers wasn't readable for this story.
| Rank | System | Country | Rmax (PFlop/s) | Power (MW) | GFlops/W* |
|---|---|---|---|---|---|
| 1 | LineShine | China | 2,198.4 | 42.2 | ~52 |
| 2 | El Capitan | US | 1,809.0 | 29.7 | ~61 |
| 3 | Frontier | US | 1,353.0 | 24.6 | ~55 |
| 4 | Aurora | US | 1,012.0 | 38.7 | ~26 |
| 5 | JUPITER Booster | Germany | 1,000.0 | 15.8 | ~63 |
*Calculated by TechMeasured from the list's Rmax and power figures.
LineShine at the National Supercomputing Centre in Shenzhen runs on 13.79 million cores of Huawei's ARMv9-based LX2 CPUs. It is about 21% faster than El Capitan on the HPL benchmark and also leads the HPCG ranking at 22 PFlop/s. On this list's own numbers, though, it does less work per watt than El Capitan, Frontier or JUPITER. MERICS analyst Antonia Hmaidi argued in September that the CPU-only design, shaped by US export controls, makes LineShine "unlikely to meaningfully help" with China's bottleneck in training AI models. She noted it placed only fourth on the AI-oriented HPL-MxP benchmark.
What to watch
- Rubin HPC's FP64 and memory specs. Until Nvidia publishes them, any comparison with AMD's MI-series or with today's exascale machines is guesswork.
- Independent Vera benchmarks. Nvidia's 2x efficiency claim uses unnamed x86 baselines, and StorageReview's own test showed a smaller lead.
- Low TOP500 turnover. A static top 10 alongside record AI spending suggests the biggest new machines are being built for private AI clusters that don't submit results to the list. That is an inference, not something TOP500 has stated.
Sources
- NVIDIA Blog — NVIDIA Announces Rubin HPC — A New GPU to Accelerate Agents for Science, Engineering and Analytics blogs.nvidia.com
- GlobeNewswire — NVIDIA Expands Vera CPU Platform to Accelerate Scientific Discovery globenewswire.com
- TOP500 — November 2026 list new.top500.org
- TOP500 — Home (68th list announcement, Nov 16, 2026) new.top500.org
- StorageReview — NVIDIA SC26: Rubin HPC GPU, Vera CPU HPC rack, Nemotron for Science, quantum decoder storagereview.com
- Unite.AI — NVIDIA Unveils Rubin HPC GPU for Agent-Driven Science and Engineering unite.ai
- MERICS — Supercomputer LineShine: A restrictions-driven leap with limitations merics.org