The team identified NVIDIA DGX SuperPOD reference architecture as the answer — specifically, a DGX SuperPOD with 248 NVIDIA Hopper GPUs designed for large-scale research access. Named Marlowe, the system launched as part of a larger Stanford compute ecosystem.

The introduction of Marlowe added a complementary, GPU-optimized cluster built on a unified DGX SuperPOD architecture — giving researchers a clear, campus-wide path for planning how to scale from initial experiments to large, multi-node jobs —  with a primary focus on large-scale AI model training while providing flexibility for fine-tuning and inference. NVIDIA Base Command Manager controls the entire cluster for simplified GPU provisioning and administration.  

Marlowe complements the diverse NVIDIA-certified partner solutions already supporting research across Stanford — from Sherlock’s large-scale, shared environment to specialized departmental clusters. By serving as a blueprint for high-performance, scalable GPU workflows, Marlowe helps teams refine their approaches and more easily extend those practices across NVIDIA-certified systems as their computational needs grow.

"Stanford's research computing ecosystem empowers world-class discovery through specialized services designed for every scale of computation,” said Ben Rogers, executive director of Stanford Research Computing. “Marlowe represents the pinnacle of this ecosystem — enabling researchers to scale across dozens of GPUs and beyond. This transformational capability accelerates breakthroughs in AI, medicine, and other fields that will help shape our future.” 

The real innovation wasn't just the infrastructure, but also the collaborative approach to deployment and operations. NVIDIA worked closely with Stanford leadership at every level, from executive leadership shaping policy, faculty champions driving technical adoption, and operations teams ensuring day-to-day success.  

Mark III Systems, an NVIDIA DGX SuperPOD Specialization Partner,  played an essential role in the deployment and ongoing support of Marlowe, ensuring smooth integration within Stanford’s broader compute ecosystem. NVIDIA's solutions architects also work closely with Stanford’s research computing teams and provide hands-on guidance to researchers. This collaborative model helps sustain long-term innovation and ensures that the benefits of Marlowe reach across disciplines and research groups.

Marlowe was built differently from traditional HPC clusters. Rather than complex allocation systems, it offers a streamlined, research-focused approach. Any principal investigator can apply for access, undergo a computational suitability review, and start with basic access to the system. As their work progresses and they are ready to scale to larger jobs, they can apply for medium and large allocations that provide consistent, non-preemptible access for multi-node training.

The NVIDIA Quantum-2 InfiniBand networking platform capabilities built into the DGX SuperPOD architecture have proven transformational for these users. Researchers can scale across hundreds of NVIDIA Hopper GPUs with the kind of high-speed interconnect that makes industry-scale training jobs feasible, turning workloads that once required careful queuing or external resources into day-to-day practice on campus.

One of the early Marlowe users is Tong Wu, a postdoctoral researcher at Stanford working with Gordon Wetzstein, professor of electrical engineering and computer science, on long-horizon video world models.  

With access to Marlowe, Wu and team were able to train significantly larger diffusion transformer–based models on substantially larger video datasets than before, dramatically reducing training time and accelerating their experimental iteration cycle. This computational scale was critical for developing the paper “Video World Models with Long-term Spatial Memory,” which introduces a geometry-grounded, long-term spatial memory that explicitly stores and retrieves 3D scene information over time. 

In the paper, the team reports that their framework improves long-term quality, consistency, and effective context length compared to strong autoregressive video baselines—maintaining scene structure and camera behavior more faithfully during revisits. Their evaluations show that incorporating the long-term spatial memory leads to higher visual quality and better environment recall across extended sequences, paving the way toward more stable, long-horizon world generation. 

Marlowe has enabled the Stanford Computational Imaging Lab to not only pursue new research directions in spatial-aware video generation and world models but also to collaborate with researchers globally who are working on related topics in controllable video generation and physical AI. With this foundation, the lab can continuously produce high-impact work at the frontier of video world modeling and share results with a fast-moving international community.

Thierry Tambe, professor of electrical engineering and computer science, worked with his research group and Tsachy Weissman’s lab to use Marlowe to scale GaussianVision: Vision-Language Alignment from Compressed Image Representations using 2D Gaussian Splatting—a project that asks whether raw, redundant RGB pixels can be replaced with a more compact, semantically rich visual substrate for large-scale vision-language models.  

Using Marlowe, the team developed a scalable 2D Gaussian splatting (2DGS) pipeline, achieving over 90x faster fitting and around 97% GPU utilization compared to prior implementations.  

Building on this foundation, the team adapted leading vision-language models (VLMs) to compact 2D Gaussian splat inputs, reducing training to under 14% of model parameters while still matching or beating strong pixel-based baselines across key benchmarks.

This work, accepted to CVPR 2026 as GaussianVision, represents the first large-scale study of 2DGS as a visual substrate for vision–language alignment and illustrates how Marlowe’s supercompute scale made thousands of experiments spanning compression, pretraining, and end-to-end VLM evaluation achievable.

Together, these projects showcase how Marlowe turns ambitious, compute-intensive ideas into repeatable research programs—enabling Stanford teams to explore entirely new model architectures and representation paradigms at production scale.

See Marlowe in Action

The impact of Marlowe extends beyond individual research projects. Stanford Data Science acts as a collection hub for top researchers across all seven schools that are encouraged to use Marlowe to grow their technical capabilities and scale their workloads. 

"Marlowe is exciting because it gives Stanford researchers frictionless access to supercomputing, enabling critical computational research that would otherwise be difficult—or even impossible—to carry out,” said Emmanuel Candès, associate director of Stanford Data Science. 

Over eighteen million GPU hours have been utilized annually across Stanford's entire compute ecosystem since Marlowe’s inception, supporting research that spans from materials science and neuroscience to medical imaging and AI—positioning Stanford as a leader in GPU-accelerated academic research. 

Today, Marlowe represents something larger than a computing cluster. It's a symbol of Stanford's commitment to removing infrastructure barriers so researchers can focus on what they do best: pushing the boundaries of human knowledge.