GPU-based machine learning is taking shape across two connected areas: building models and running them. A tutorial published on September 12, 2026, by Sana Hassan examines that process through NVIDIA’s cuML and the RAPIDS ecosystem, while Nvidia and d-Matrix have announced a roadmap focused on fast inference.
Together, these developments show a broad view of machine learning infrastructure. The tutorial focuses on implementation, benchmarking, explainability, clustering, and model inference. The Nvidia and d-Matrix collaboration focuses on inference, creating a second point of attention around how trained models are run.
What the cuML tutorial covers
Hassan’s tutorial begins with configuring a GPU environment and accelerating scikit-learn workloads with cuml.accel. That gives the walkthrough a practical starting point: it connects familiar scikit-learn workloads with NVIDIA’s GPU machine learning framework instead of treating model development and hardware as separate subjects.
The tutorial then benchmarks CPU and GPU implementations across several machine learning tasks. These include principal component analysis, or PCA, K-Means, nearest-neighbor search, logistic regression, random forests, and DBSCAN. The range matters because it covers dimensionality reduction, clustering, search, classification, and ensemble models within one workflow.
Benchmarking is only one part of the tutorial. It also builds GPU-based manifold-learning and clustering pipelines with UMAP, t-SNE, and HDBSCAN, then uses trustworthiness metrics to examine those pipelines. These tools bring together data representation, cluster discovery, and a way to evaluate the resulting manifold-learning work.
The tutorial also explores forest inference with FIL, focusing on high-throughput use of trained forest models. It validates GPU-generated SHAP explanations, which connects model inference with explainability. That combination gives the walkthrough a clear focus on both producing predictions and examining the explanations generated for those predictions.
Other parts of the tutorial cover hyperparameter optimization with scikit-learn meta-estimators and the serialization of trained models. In practice, that means the material spans model configuration, evaluation, explanation, optimization, and saving trained models for later use.
From model workflows to fast inference
The Nvidia and d-Matrix announcement adds a separate but related piece to the picture: a roadmap for fast inference. The collaboration places inference alongside the training and evaluation workflows covered in the cuML tutorial.
That distinction helps clarify the roles described in the material. The tutorial follows the machine learning workflow from GPU environment setup through benchmarking, clustering, explainability, optimization, and model serialization. The Nvidia and d-Matrix roadmap addresses fast inference, the stage where trained models are used to produce results.
The two areas meet around model deployment and performance. The tutorial includes FIL for forest inference, GPU-generated SHAP explanations, and serialized models, while the Nvidia and d-Matrix announcement focuses on fast inference as a roadmap. Neither replaces the other; they describe different parts of the same machine learning path.
A heterogeneous future for AI infrastructure
The Forbes article about the future of AI infrastructure was published on September 8, 2026, at 07:45am EDT. Ido Susan, the CEO and Co-founder mentioned in the article, authored the piece as a member of the Forbes Councils.
Its central theme is that the future of AI infrastructure will be heterogeneous. That idea fits the technical range in the cuML tutorial, which moves across GPU environments, CPU and GPU benchmarks, manifold learning, clustering, forest inference, explainability, hyperparameter optimization, and model serialization.
It also connects with the Nvidia and d-Matrix roadmap for fast inference. The combined picture is not limited to one model or one stage of development. It includes the hardware environment, the software framework, the algorithms being tested, the metrics used to evaluate them, the explanations attached to predictions, and the systems that run trained models.
For readers following machine learning infrastructure, these developments offer two clear reference points. The cuML tutorial shows how a GPU-centered workflow can be implemented and examined across many tasks. The Nvidia and d-Matrix announcement shows fast inference as another infrastructure priority, while Susan’s Forbes article frames the wider direction as heterogeneous AI infrastructure.
That makes the material useful as a map of the field’s moving parts. It covers how models are built, benchmarked, explained, optimized, stored, and run, while placing those activities inside a future shaped by different forms of AI infrastructure working together.
Based on
