Samsung HBM4E, Kioxia And HPE Storage, And Peak:AIO Metadata Server

Samsung is already sampling the next generation of high-bandwidth memory—HBM4E—months before its expected commercial launch. This signals an accelerated race among memory makers to supply the insatiable appetite of AI accelerators.

Samsung Electronics has begun providing HBM4E samples to key customers, advancing the standard that promises to double the bandwidth of current HBM3E. Industry sources confirm that the new memory modules are being tested in AI training clusters and high-performance computing systems. The move comes as Nvidia, AMD, and other chip designers demand ever-faster memory to feed their graphics processing units (GPUs). Samsung aims to start volume production in the first half of next year, leapfrogging competitors SK Hynix and Micron.

High-bandwidth memory stacks DRAM dies vertically, using through-silicon vias to achieve enormous data transfer rates. HBM4E is expected to deliver up to 1.2 TB/s per stack, up from around 1 TB/s in HBM3E. Each stack may pack 48 GB of capacity, double the current maximum, thanks to 6th-generation 1d DRAM technology. The samples are being validated by major data center operators and AI companies, who are already designing next-generation server platforms around the new memory.

In a separate development, Kioxia has teamed up with Hewlett Packard Enterprise to optimize storage solutions for both terrestrial and space-based computing. Kioxia’s Flash Memory and Storage Class Memory will be integrated into HPE’s ProLiant servers and space-qualified systems. The collaboration targets low-latency, high-endurance storage for edge AI and satellite data processing. Kioxia is also exploring 3D NAND radiation hardening for orbital environments.

Meanwhile, Peak:AIO and Los Alamos National Laboratory have jointly developed a metadata server that breaks the storage bottleneck in large-scale AI workloads. Traditional file systems struggle with the massive metadata operations generated by training datasets. The new Peak:AIO Metadata Server uses compute express link (CXL) and a distributed architecture to handle billions of metadata operations per second. In testing, it reduced data staging time by 70% and allowed 80% more GPU utilization during training cycles.

These three announcements underscore an industry-wide effort to eliminate I/O bottlenecks that limit AI performance. Memory bandwidth, storage latency, and metadata management are becoming as critical as raw compute power. “We are hitting a wall where memory and storage are the constraining factors, not the processors themselves,” said Dr. Lisa Su, CEO of AMD, at a recent conference. “Innovations like HBM4E and advanced storage-class memory are essential to sustain the pace of AI progress.”

Looking ahead, Samsung’s HBM4E sampling gives it a potential first-mover advantage in the next memory cycle. Kioxia and HPE plan to demonstrate their space storage platform at the 2027 International Space Station Research and Development Conference. Peak:AIO expects to commercialize its metadata server in the second half of 2026. The convergence of faster memory, smarter storage, and metadata acceleration will define the next generation of AI infrastructure.