Marvell Technology and Penguin Solutions announced a collaboration to develop a Photonic Memory Fabric based on Celestial AI’s “HBM at a distance” concept. Marvell will supply the core modules, while Penguin will build the system hardware, power management, software and management layer; Marvell expects the first-generation expansion system to launch in the second half of 2027. Vendor-claimed specifications include 72 GB of HBM3e and eight-channel DDR5 support per module, with the first chassis accommodating 16 modules.
openEuler previewed planned Agentic Infra technical work and supernode deployment progress at a media briefing for Agentic Infra OS Conference & openEuler Summit 2026. The reported areas include agent observability and security governance, fast-starting sandboxes, agent memory systems, and KV Cache optimization, with openEuler also describing OS-level improvements to filesystem, memory management, security, and resource scheduling. In an advertising business example, openEuler claimed that optimizing data transfer between DDR memory and accelerator cards reduced KV Cache read/write latency by nearly 84%.
JD Advertising and Huawei Kunpeng deployed a Kunpeng supernode architecture for generative advertising inference, using UnifiedBus to connect larger memory pools, two-level KV caching and revised RPC communication. The partners reported that the system cut end-to-end inference latency by 15% and handled 5% of traffic in core advertising scenarios during the 618 peak. These performance figures come from the collaborating parties, and the deployment does not yet cover all search, advertising and recommendation workloads.
SK hynix development chief Ahn Hyun said in a keynote at SEDEX 2026 in Seoul on October 14 that AI infrastructure performance scale will grow about 20 times by 2030, according to Yonhap. He said the AI inference market is expanding along three axes — task complexity, context length and request volume — swelling the KV cache and forcing data to be spread across HBM, DRAM and SSD, with the resulting data movement becoming a new bottleneck on performance and power. SK hynix said it is developing 3D stacked DRAM that places DRAM directly atop logic chips such as GPUs and NPUs, plus high-bandwidth flash (HBF), which applies HBM's through-silicon-via stacking and high-speed I/O to NAND; it aims for HBF bandwidth at about 50% of same-generation HBM with roughly 10x the capacity, targets that have no independent measurement yet.
Korean outlet ZDNET Korea reports that Samsung Electronics is developing two NAND flash products tuned for AI workloads, zNAND-T and zNAND-I. zNAND-T stands for throughput, using TSV-stacked NAND dies to optimize read/write bandwidth for data-center AI silicon, an approach the report likens to the HBF concept proposed by SanDisk and SK hynix; zNAND-I refers to IOPS and targets random performance in scenarios with frequent small reads and writes, such as KV cache offloading. Separately, Samsung announced a next-generation memory and storage technology strategy for AI infrastructure spanning high-performance memory, Compute Express Link (CXL)-based storage expansion, and new-generation SSD data storage and management. Together with the previously disclosed zNAND-O, the three zNAND variants are positioned for different workloads, which the report frames as a shift from general-purpose NAND toward customization.
What problem it solves
Developers often need to understand why model applications are slow or costly, but system decisions require reasoning across model architecture, hardware limits, memory, computation, communication, and scheduling. Easy-to-miss factors such as KV cache, bandwidth limits, and inter-device communication can shift estimates by orders of magnitude, so system design should begin f…
Handing a problem to four Agents most easily produces four attempts. Increasing the number to eight may raise the success rate further. But if every Agent searches along similar lines, a significant portion of the additional computational power brought by a larger team will be spent on duplicate work. If we instead require them to communicate continuously, messages will consume context and reasoni…
What problem it solves
Running frontier-scale open-weight models locally requires an inference platform that can make efficient use of the heterogeneous hardware available on a personal device. The project addresses this by treating GPUs, CPUs, host memory, and interconnects as a unified, elastic inference platform.
Key features
Edge-native MoE runtime — The runtime provides MoE serving with bandw…