Coming online in Q1 2027
Kitsune's hardware is arriving soon. After an initial testing phase, Kitsune will be allocated through the NSF ACCESS program. Researchers who want launch updates can sign up. ACCESS also hosts the Kitsune Affinity Group, its community group for cluster users.
Kitsune is the flagship research cluster of the Stony Brook Supercomputing Center, the university's new Tier 1 supercomputing center, funded by the National Science Foundation through its Advanced Computing Systems and Services program (award OAC-2505376, Sustainable Cyber-infrastructure for Expanding Participation) and operated in partnership with the University at Buffalo's Center for Computational Research. It is built, like everything here, for the long tail of science: modern computation, accessible to research groups across the university. The project was first announced under the working name AMA27; Kitsune is its name now, though the award title keeps the original.
Why "Kitsune"? Kitsune (狐) is Japanese for fox. The name continues a tradition begun with Ookami (狼, wolf), SBSC's previous NSF-funded ACCESS resource, and tips its ears to the Sea Wolf, Stony Brook's mascot. The fox's prominent tail is the point: Kitsune is designed for the long tail of scientific research, the many workloads that never make headlines but make science run.
Timeline
- Proposal submitted: 2024, NSF Advanced Computing Systems and Services solicitation 24-583 (Category I, capacity resources)
- Award funded: NSF OAC-2505376, July 2025
- Order in place: October 2026
- Hardware arriving: soon
- Benchmarking and early testing: to follow arrival
- Comes online: Q1 2027
- ACCESS allocation: after the initial testing phase
- General availability: to be announced
At a glance
| Item | Status |
|---|---|
| Hardware | Arriving soon; commissioning and benchmarking to follow |
| Scale | 162 nodes · more than 60,000 AMD cores |
| Comes online | Q1 2027 |
| Allocation | NSF ACCESS, after the initial testing phase |
| Documentation | Opens as Kitsune comes online; current docs cover SeaWulf |
| Stay in touch | Launch-update sign-up |
Planned specifications
| Item | Value |
|---|---|
| Compute nodes | 162 |
| Peak performance | 6.7 PFLOPS DP |
| Memory | 126 TB |
| Storage | 6 PB usable GPFS |
| Interconnect | InfiniBand NDR-400 |
| Operating system | Rocky Linux 10 |
| Item | Value |
|---|---|
| Processor | 2× AMD EPYC 9965 (Turin, Zen 5) |
| Cores per node | 384 (192 per socket) |
| Peak per node | 41.2 TFLOPS DP |
| Memory per node | 768–1500 GB, 24 channels at 6400 MT/s |
| Clock (base / all-core / single-core max) | 2.25 / 3.25 / 3.7 GHz |
| Vector | AVX-512, full 512-bit data path, with VNNI and BF16 |
| Cache | 48 KB L1d / 32 KB L1i and 1 MB L2 per core; 384 MB L3 per socket |
Why this system? Versatile HPC for scalar, vectorized, and memory-intensive workloads, with high throughput: many independent jobs at once alongside large parallel jobs. It also handles CPU-based AI inference for local LLM workflows, with strong energy efficiency and useful performance per dollar.
Who Kitsune is for
- High-throughput work: many independent jobs running at once
- Individual jobs that scale across many cores; multithreaded and MPI applications
- Scalar and vectorized computation, and jobs needing substantial memory capacity
- CPU-based AI inference and text-processing / NLP preprocessing
Examples include read mapping and other bioinformatics pipelines, scientific simulation and modeling, and parameter sweeps and batches of independent simulations, with many users at once.
Kitsune and SeaWulf
| Feature | SeaWulf | Kitsune |
|---|---|---|
| Status | In production | Hardware arriving soon; comes online Q1 2027 |
| Operated by | Research Computing and Informatics | SBSC with the University at Buffalo's CCR |
| Processors | SeaWulf III: 94 compute nodes Intel Sapphire Rapids w/ HBM, 48 AMD Milan; SeaWulf II: 164 Intel Haswell, 64 Intel Skylake | 162 nodes, each 2× AMD EPYC 9965 (Turin), for more than 60,000 cores |
| Accelerators | 11 GPU nodes, each 4× A100 80 GB (SeaWulf II: 28 K80, 2 P100, 2 V100) | CPU-only by design, built for CPU-based AI inference |
| Available to | Campus and Brookhaven Lab staff (request an account) | Launch-update sign-up below; ACCESS allocation after initial testing |
| Documentation | SBSC docs | Opens as Kitsune comes online |
Benchmarking Kitsune
First results will appear here once the hardware is in and early testing begins. Benchmarks will come first, and early-testing visualizations will live in this section.
Launch updates
Get launch updates
Want updates as Kitsune comes online, including first results, documentation, and how to get time on the system? Sign up with the Kitsune interest form (or email [TEAM ADDRESS, PENDING]) with your name, department, and what you would run on Kitsune, and we will add you to the list. New to ACCESS? Getting started there now will make the allocation path smoother.