computer memory

An educational post explains that modern computers rely on a hierarchy of memory and storage technologies — from fast, small caches to large, slower storage — to balance speed, capacity, and cost. It notes that not all computer memory performs equally, and links to freely downloadable PDF ebooks containing related technology infographics for learners.

A question about why computers and phones offer a separate restart option in addition to power off is drawing discussion. The common explanation: restart reboots the system automatically, clearing memory and refreshing software without needing someone to press the power button again, which matters for updates, fixing glitches, and devices like phones where powering back on manually is inconvenient.

Industry commentary highlights that faster AI processors from Nvidia and AMD cannot deliver gains if high bandwidth memory (HBM) cannot feed data quickly enough. The discussion frames memory bandwidth, not raw compute, as a key bottleneck for AI data centres, keeping HBM supply and performance central to semiconductor debates.

SK Hynix has introduced MARS, a Memory-Centric AI Rack System, at the Open Compute Project Global Summit. The system is aimed at data center infrastructure, reportedly leveraging CXL technology to expand memory capacity alongside GPUs for AI workloads. The announcement underscores the chipmaker's push into memory-driven server architecture for AI data centers.

Broadcom chief executive Hock Tan says the next constraint on artificial intelligence growth will not be chips, power, or memory, but something outside technology altogether. His remarks, reported by financial media, suggest the limitation lies with adoption, demand, or economic factors rather than hardware supply. The comments are drawing attention as companies worldwide race to secure computing capacity for AI and debate where the real limits to expansion lie.

Reexen Technology has raised $150 million to develop its compute-in-memory AI chips, a chip architecture that processes data directly in memory to cut the power and speed bottlenecks of traditional designs. The funding positions the company among a wave of startups competing to supply specialised hardware for the booming artificial intelligence market, where demand for efficient chips keeps outstripping supply.

Micron Technology has laid out a product roadmap detailing its plans for high-bandwidth memory (HBM), DRAM and NAND as demand for AI computing accelerates. The company is positioning itself as a key memory supplier for data centers and AI hardware, with HBM a central focus as chipmakers compete to feed the memory needs of AI accelerators.

Investors are watching CXMT along with two other AI chip stocks as interest in semiconductor companies tied to artificial intelligence demand continues. CXMT, a Chinese memory chip maker, is drawing attention amid growth in memory needs for AI computing. The report flags these names as worth monitoring for investors tracking the AI hardware sector.

Microcontroller vendors are adding multimodal compute and tiered memory to edge AI chips, allowing AI models once confined to the cloud to run on traditional MCU platforms. The development, reported by EE Times, signals a push to bring smarter inference directly into embedded and low-power devices without constant cloud connectivity.

Perplexity has released Perplexity Computer, an AI agent that runs entirely on Apple silicon Macs, keeping files on the device. It is available to Pro, Max and Enterprise subscribers, with the company recommending 36GB of RAM to run it. The local-first approach is drawing attention for its privacy implications and its steep memory requirements.

AI inference at scale is increasingly constrained by memory capacity and cost rather than compute. Huawei is promoting its OceanStor M900, a memory storage platform aimed at improving the economics of large-scale inference workloads. The Register reports the company positions the product as a way to ease the memory bottleneck limiting AI deployment growth.

A technical article on one-hash Bloom filters, a space-efficient probabilistic data structure variant that uses a single hash function instead of the conventional several, is being discussed on Hacker News. Commenters are weighing the trade-off between fewer hash computations and the higher false-positive rate, with discussion touching on memory efficiency and practical use cases.

Yale University researchers have developed a fluidic memory chip that stores information using water flow through microfibers instead of conventional solid-state electronics. The approach points to new possibilities for computing in wet or biological environments, where traditional silicon chips struggle. Details on performance and commercial applications remain limited as the work draws early attention.