Serving large numbers of concurrent users across modern AI inference workloads – multi-turn chatbots and long-context agentic reasoning – is fundamentally a memory capacity problem. The High Bandwidth Memory (HBM) on each GPU must be shared between static model weights and the growing key-value (KV) cache of every active session.
As user concurrency scales, the KV cache exhausts the HBM headroom that remains after model weights are loaded, causing latency to spike long before the GPU's compute capacity is reached. This paper describes how the Samsung Cognos AI Memory Agent and the SGLang HiCache hierarchical KV cache layer, working together, solve this bottleneck at the infrastructure level.
Details
As data-intensive workloads continue to grow in scale and complexity, the traditional reliance on expensive, high-speed DRAM creates a critical imbalance between memory capacity, cost, and performance. This issue leads to underutilized CPUs and high total cost of ownership (TCO) in modern data centers. A tiered memory architecture addresses this challenge by using DRAM as a small, fast caching layer on top of a large cost-effective, highly dense CXL or NVMe memory.
We introduce Cognos — a memory software orchestrator that abstracts the memory tiering based on the user application SLAs.
Details
Samsung, being the world leader of DRAM and Flash memory technologies, has recognized early on the potential for a CXL based solution architecture. Samsung heavily invested into building in-house technologies, know-how, and device solutions to fully enable the capabilities that CXL brings to the next-generation system architecture. These investments include the CXL Memory Module – DRAM (CMM-D) device and the solution level device CXL Memory Module – Memory Box (CMM-B).
Additionally, Samsung has invested in a set of world-class firmware, drivers, APIs and management software to enable customers to quickly adopt these solutions that are aimed at reducing their total cost of ownership.
Samsung Cognos provides the glue that stitches all of the CXL-based technologies and devices into one homogeneous and integrated solution that developers and customers can easily adopt and use in their cloud or private data center solutions and services. Cognos includes a software developer kit that provides the foundational software components that are needed to quickly utilize the new CXL-based devices with customer applications and a host of management software, including
a GUI to manage one-to-many server/switch nodes to handle rack-level cluster/system management. Samsung Cognos addresses memory stranding problems by pooling memory into a global resource that is shareable and reusable by resources across the data center. It provides a value-added software layer for Memory Box and CMM-D in the host to enable easier adoption by applications with easy-to-use interfaces.
Details
Zero-ETL is an approach that aims to minimize or eliminate the traditional ETL processes that are typically used to extract data from source systems, transform it into the desired format, and then load it into a data warehouse or data lake for analysis. In the context of Zero ETL, the data is often accessed in real-time or near real-time, eliminating the need for batch processing. This approach is becoming more popular with the advent of technologies like Change Data Capture
(CDC) and streaming data pipelines.
Details
Optimizing Data Architecture with Samsung SVK & VMware Greenplum
Ivan Novick, Director, VMware & David McIntyre, Director, Samsung & Pramod Peethambaran, Director, Engineering, Samsung
About this talk :
Join our informative webinar for a detailed review of a data architecture, crafted with Samsung’s advanced storage solutions and VMware Greenplum’s data processing capabilities. We’ll walk through the design principles, key architectural elements, and their role in enhancing system performance. We’ll also present test data to provide a clear picture of the architecture’s performance and explain the trade-offs made during the design process. This
session aims to equip you with practical insights to leverage the latest Samsung hardware technologies and build high-performance systems. Immerse yourself in this exploration of data warehousing and analytics optimization.
Video
High Performance Object Storage (HPOS) stack enables disaggregated storage over multiple CMM-HC (Samsung’s computational storage drives). A data centric solution abstracting which computational function runs in which accelerator layer (i.e. CPU, GPU, XPU) with ease of scale. Samsung’s HPOS solution provides reference solution for Video AI applications requiring in-storage video pre-processing. And for Data Analytics use case providing high performance for unstructured/object
data in data lake.
Video
Posted whitepaper in 2023 MemCon
Increasingly, as AI technology evolves into more sophisticated applications, training dataset sizes continue to grow exponentially. In order to scale storage and network infrastructure commensurately to deliver the data required and avoid unbearable training cycle times, there is a need for a new storage concept and innovative solution to address this technology gap. DSS Gen2 and beyond provides a potential next generation architecture and solution to alleviate this
bottleneck.
Another issue that is major impediment to data center scaling is power utilization. And since one of the key storage consumer is data storage, DSS technology not only plan to optimize server power utilization but also partner and leverage Samsung’s new high capacity SSDs which has one of the highest density in the world.
Full Whitepaper
Github
2022 OCP Global Summit
With the advent of new application workloads related to Big Data including AI/ML, IoT, Video, security and many other machine generated data, there is a strong incentive for companies as well as governments to mine this treasure trove of data to extract value. Innovative companies taking on these storage challenges have tried storing Big Data into data lakes and moving them into locally attached storage for analysis but due to the sheer size of the data set this turns out to
be too time consuming. Traditional network data storage also have been explored but overcoming inherent scaling and performance bottlenecks are difficult. DSS provides an innovative solution to the problem by designing purpose built storage that only targets these specific workloads.
Video
Slide
Global
