The National Institute of Health's modernized high-performance computing (HPC) capabilities enable significant innovations in scientific research. NIH's supercomputer Biowulf is the first and largest supercomputer completely dedicated to advancing biomedical research, and produced the first complete, gapless sequence of a human genome. Looking to the future, HPC's emerging scientific requirements include compatibility with petabyte-scale data, increased CPU computing, additional storage capacity, and flexibility. To face these challenges, HPC is exploring transformative technologies such as Gen AI and quantum computing to help improve the discovery of scientific research. This presentation will discuss the current state of HPC, limitations within HPC, and future exploration of Gen AI and quantum.

Compute Express Link (CXL) has risen as a promising interconnect technology that enables seamless high-speed, low-latency communication between host processors and various peripheral devices, making it attractive for memory-intensive applications. In this talk, we'll evaluate the performance characteristics of CXL memory in real-world environments using ASIC-based CXL memory. We then present our learnings from microbenchmarks, explore the potential use cases and evaluate their benefits from using CXL memory. Based on our comprehensive evaluations, we share our insights of how we may better utilize CXL memory to optimize the performance and cost of data center applications, as well as our vision for next-gen memory architecture leveraging new capabilities from CXL 2.0/3.x.

Online commercial app marketplaces serve millions of apps to billions of users in an efficient manner. Bandit optimization algorithms are used to ensure that the recommendations are relevant, and converge to the best performing content over time. However, directly applying bandits to real-world systems, where the catalog of items is dynamic and continuously refreshed, is not straightforward. One of the challenges we face is the non-trivial computation costs for large scale systems, which is further aggravated by user privacy related constraints for server side computation. To address this problem we introduce an efficient two-layer bandit approach which is contextualized to user cohorts of similar taste. We mitigate cannibalization at runtime within a single multi-intent content surfacing platform by formalizing relevant offline evaluation metrics, and by involving the cross-component interactions in the bandit rewards. The framework allows flexibility for tradeoffs between compute, storage and accuracy of  models. The user engagement in our proposed system has more than doubled as measured by online A/B testings.

The presentation delves into the evolution, current state, and prospective developments within data-driven machine learning. In an era where data has ascended to the status of a pivotal resource, this presentation emphasizes its indispensable role in shaping the landscape of machine learning and how these changes have significantly influenced systems infrastructure.
Delving into the past, it meticulously examines the historical origins of data-driven modeling, charting its progression from rudimentary concepts to the intricate algorithms that underpin modern machine learning. The presentation illuminates early techniques like perceptrons and decision trees and elucidates their enduring impact on the field.
In the present, this presentation expounds upon the transformative influence of big data and deep learning, illuminating real-world applications while highlighting the associated challenges and opportunities that have engendered profound alterations in systems infrastructure.
As we look towards the future, this presentation provides invaluable insights into emerging trends and technologies such as quantum computing and edge AI, poised to redefine the future of machine learning and further revolutionize systems infrastructure.
By amalgamating theoretical insights, empirical observations, and forward-looking perspectives, this presentation offers a comprehensive overview of the past achievements, current dynamics, and potential future scenarios in the realm of data-driven machine learning, shedding light on how these changes have reshaped systems infrastructure.

Oracle AI Vector Search enables enterprises to leverage their own business data to build cutting-edge generative AI solutions. AI Vectors are data structures that encode the key features or essence of unstructured entities such as images or documents. The more similar two entities are, the shorter the mathematical distance between their corresponding AI vectors. With AI Vector search, Oracle Database is introducing a new vector datatype, new vector indexes (in-memory neighbor graph indexes and neighbor partitioned indexes), and new Vector SQL operators for highly efficient and powerful similarity search queries. Oracle AI Vector Search enables applications to combine their business data with large language models (LLMs) using a technique called Retrieval Augmentation Generation (RAG), to deliver amazingly accurate responses to natural language questions. With AI Vector Search in Oracle Database, users can easily build AI applications that combine relational searches with similarity search, without requiring data movement to a separate vector database, and without any loss of security, data integrity, consistency, or performance. At the heart of this are the hardware and system requirements needed to facilitate scale up and scale out AI vector search.

Compute performance demand has been growing exponentially in recent years, and with the advent of Generative AI, this demand is growing even faster. Moore’s law coming to an end as well as the Memory Wall (bandwidth & capacity) are the main performance bottlenecks. The chiplet system-in-package (SiP) is the industry's solution to these bottlenecks. Silicon interposers are industry’s main technology to connect chiplets in SiPs, but they introduce several new bottlenecks. The largest interposer going to production is 2700mm2, which is ~1/4 the largest standard package substrate. Thus, a SiP with silicon interposer has limited compute & memory chiplets, thus limited performance.
This presentation introduces Universal Memory Interface (UMI), a high bandwidth efficient D2D connectivity technology between compute and memory chiplets. UMI PHY on standard packaging provides similar bandwidth/power to D2D PHYs with silicon interposers, thus enables creation of large & powerful SiPs required to address Gen AI applications.

Shell Upstream has been processing large subsurface datasets for multiple decades driving significant business value.  Many of the state of the art algorithms for this have been developed using deep domain knowledge and have benefitted from the hardware technology improvements over the years. However, the demand for more efficient processing as datasets get bigger and the algorithms become even more complex is ever-growing. This talk will focus on the memory and data management challenges for a variety of traditional HPC workflows in the energy industry. It will also cover unique challenges for accelerating modern AI-based workflows requiring new innovations.