Catching up on some recent storage industry announcements from Samsung, Kioxia and Peak:AIO and Los Alamos Laboratories.
Samsung said that it has begun shipping 12-layer HBM4E samples to major global customers. Earlier this year the company started mass production of its HBM4 product earlier this year. High Bandwidth Memory, HBM, stacks DRAM die on top of each other with massive parallel connectivity to maximize the data bandwidth. HBM memory is used in conjunction with GPUs and other domain specific processors in a chiplet architecture, storing data to support AI training and inference workloads. The front and back of the HBM4e HBMs are shown below.
Currently HBM memory is in high demand with not enough production to meet that demand with consequently high prices and with big users such as hyperscale data centers having long term contracts with suppliers to get the HBM they need.
According to the company, Samsung’s HBM4E delivers a stable pin speed of 14 gigabits-per-second (Gbps), with performance scalable up to 16Gbps to support increasingly intensive data processing requirements. This represents more than a 20% increase over its HBM4, while delivering memory bandwidth of up to 3.6 terabytes-per-second (TB/s) per stack, helping maximize computing performance for large language models (LLMs) and next-generation AI systems.
Samsung’s 12-layer HBM4E is offered in a 48-gigabyte (GB) capacity, representing more than a 30% increase over the previous generation, with plans to expand the lineup to include 32GB (8-layer) and 64GB (16-layer) configurations in accordance with customer requirements.
At the HPE Discover conference Kioxia will highlight the use of their SSD products in terrestrial and space-based computing.
Kioxia will be featuring the company’s PCIe 5.0 Enterprise and Datacenter standard form factor, EDSFF and 24G SAS, SAS-4, products. These products provide higher performance, increased storage density and improved power efficiency over previous generation products. Kioxia SSDs are used in a broad range of HPE platform spanning enterprise servers, digital storage, mobile computing and modern data center environments.
Kioxia will be highlighting its contributions to HPE’s Spaceborne Computing missions, including projects aboard the International Space Station, ISS, and future lunar exploration efforts. We have written previously about HPE servers on board the ISS using Kioxia SSDs. We have also written recently about the great number of projects to create data centers in space.
At the recent IEEE International Conference on Massive Storage Systems and Technology, MSST, Peak:AIO and Los Alamos National Laboratory launched Lattice, an open-source pNFS metadata server. Lattice introduces a new distributed metadata architecture designed to eliminate one of the most persistent constraints in large-scale AI and high-performance computing infrastructure.
According to Peak:AIO, AI infrastructure demands reshape the storage market. AI workloads, including training large models, running inference at scale and serving agentic AI applications, require ultrafast, parallel access to massive datasets continuously and reliably. GPU compute has scaled dramatically, but the storage layer feeding it, and particularly the metadata architecture coordinating it, has not kept pace.
Average GPU utilization across 23,000 production clusters is 5% (Cast AI), not because the hardware is inadequate, but because the software feeding it cannot keep up. The metadata bottleneck in parallel storage systems has become a critical constraint on AI workload performance.
Lattice resolves this architectural limitation. Built as a Linux-based, user-space pNFS metadata server designed for scale, modularity and distributed coordination. Open-source, community-supported and launched under the Linux Foundation, Lattice separates the metadata control plane into four distinct layers: Protocol State Plane, Lattice Core, MD Catalog Authority and Data Server Control Plane.
This architecture makes metadata services truly elastic for the first time, allowing them to spin up dynamically on commodity hardware whenever and wherever needed, from a single server to more than 1,000 metadata servers.
Samsung is sampling HBM4E samples. Kioxia works with HPE on terrestrial and space computing. Peak:AIO works with Los Alamos Laboratory to break metadata bottleneck in storage systems.
