# memory Published articles for memory. This is one page of public article previews, not the complete archive. Follow Next page to continue. Summaries are not the original full articles. ## How tetrahedral cages significantly reduce BVH memory usage DevFeed: [How tetrahedral cages significantly reduce BVH memory usage]() Original publisher: [Read original article]() Author: Holger Gruen Published: 2026-09-17T13:00:00Z Content type: tutorial Language: en Sources: [AMD GPUOpen]() Topics: [GPU](), [Graphics](), [Simulation and Design]() Tags: [animation](), [article-release](), [directx](), [fps](), [game-development](), [graphics](), [graphics-apis](), [memory](), [microsoft-directx](), [optimization](), [performance](), [ray-tracing](), [raytracing](), [reduction](), [research](), [technical-article](), [technical-articles](), [white-paper]() ### AI overview This article explains how tetrahedral cages reduce BVH memory usage and update costs when ray tracing independently animated geometry. In an AMD sample with animated plants, the technique uses about 1.7 GB of BVH memory and approximately 3.3 ms for BVH updates per frame, compared with up to 80 GB and more than 300 ms for conventional dense triangle BLAS updates on a Radeon RX 9070 XT GPU. ### Source excerpt The AMD tetrahedral cage technique ray traces 25,000 independently animated plants at 60+ FPS while slashing BVH memory and update costs for massive dynamic scenes. ## Shared Selective Persistent Memory for Agentic LLM Systems DevFeed: [Shared Selective Persistent Memory for Agentic LLM Systems]() Original publisher: [Read original article]() Published: 2026-09-16T00:00:00Z Content type: article Language: en Sources: [Apple Machine Learning Research]() Topics: [Large Language Model](), [agent observability](), [Framework](), [Code](), [configuration](), [Access Control](), [Git](), [Model Context Protocol (MCP)](), [CSV](), [SQL](), [Grafana]() Tags: [access-control](), [agentic](), [agents](), [code](), [configuration](), [csv](), [dashboards](), [git](), [llm](), [mcp](), [memory](), [platform](), [replication](), [sql]() ### AI overview This research introduces shared selective persistent memory for agentic LLM systems. The architecture retains reusable task specifications, data schemas, tool configurations, and output constraints while discarding session-specific reasoning traces. Shared workspaces support role-based collaborative reuse, and experiments report higher task completion than no memory or full-history persistence, along with zero-token data refresh and lower token costs. ### Source excerpt Agentic LLM systems that generate code through multi-turn tool use face a fundamental context problem: each session starts from zero, discarding the configuration choices, domain constraints, data schemas, and tool-use patterns that made previous sessions productive. Naively persisting entire conversation histories is both token-inefficient and counterproductive--irrelevant context degrades generation quality. We introduce shared selective persistent memory, a memory architecture for agentic systems that identifies and retains four categories of reusable context--task specifications, data... ## Micron Shows off 512GB DDR5 RDIMM: 12TB per Dual-Socket Server at 9,200 MT/s, Volume Production in 2H 2027 DevFeed: [Micron Shows off 512GB DDR5 RDIMM: 12TB per Dual-Socket Server at 9,200 MT/s, Volume Production in 2H 2027]() Original publisher: [Read original article]() Author: Brian Beeler Published: 2026-09-15T20:18:17Z Content type: news Language: en Sources: [StorageReview.com]() Topics: [ddr5](), [servers](), [intel](), [cpu]() Tags: [capacity](), [ddr5](), [enterprise](), [generation](), [intel](), [memory](), [modules](), [performance](), [production](), [release](), [server](), [speed](), [volume]() ### AI overview Micron demonstrated a 512GB DDR5 RDIMM rated for up to 9,200 MT/s. The module can provide 12TB of memory in a 24-slot dual-socket server, with volume production scheduled for the second half of 2027. AMD and Intel are validating it for next-generation server platforms. ### Source excerpt Micron has demonstrated a 512GB DDR5 RDIMM running on multiple server platforms, which it calls the world's first module at that capacity, and says AMD and Intel are both validating it for their next-generation server platforms. The module is rated for speeds up to 9,200 MT/s, and in a 24-slot dual-socket server it puts 12TB The post Micron Shows off 512GB DDR5 RDIMM: 12TB per Dual-Socket Server at 9,200 MT/s, Volume Production in 2H 2027 appeared first on StorageReview.com. ## Astera Labs Releases Leo 2 CXL Memory Controllers and Leo X Controller for Rackscale Fabric-Attached Memory DevFeed: [Astera Labs Releases Leo 2 CXL Memory Controllers and Leo X Controller for Rackscale Fabric-Attached Memory]() Original publisher: [Read original article]() Author: Ryan Smith Published: 2026-09-15T17:00:48Z Content type: article Language: en Sources: [ServeTheHome]() Topics: [pcie](), [systems](), [datacenter]() Tags: [astera-labs](), [cxl](), [data-center](), [ddr4](), [ddr5](), [leo](), [memory](), [other-components](), [pcie](), [releases](), [server]() ### AI overview Astera Labs is launching Leo 2 CXL smart memory controllers with DDR4 support and CXL 3.2/PCIe Gen6 connectivity, alongside the Leo-X controller for fabric-attached memory in rack-scale systems. ### Source excerpt Astera Labs is launching a new generation of Leo smart memory controllers. The Leo 2 series adds support for CXL 3.2 and PCIe Gen6, while the ambitious Leo X brings the ability to attach memory expanders directly to the fabric networks of AI accelerators The post Astera Labs Releases Leo 2 CXL Memory Controllers and Leo X Controller for Rackscale Fabric-Attached Memory appeared first on ServeTheHome. ## Dense vs. MoE Models: Active Parameters, Throughput, and When to Choose Each DevFeed: [Dense vs. MoE Models: Active Parameters, Throughput, and When to Choose Each]() Original publisher: [Read original article]() Author: Elizabeth Goodman Published: 2026-09-15T17:00:11Z Content type: article Language: en Sources: [NVIDIA Developer](), [NVIDIA Technical Blog]() Topics: [Mixture of Experts (MoE)](), [Nemotron](), [Routing (disambiguation)]() Tags: [agentic-ai-generative-ai](), [ai-inference](), [architecture](), [cost](), [data-center-cloud](), [deployment](), [llms](), [memory](), [mixture-of-experts-moe](), [models](), [moe](), [nemo](), [nemotron](), [performance](), [router](), [routing](), [tokens]() ### AI overview This article explains how dense and Mixture-of-Experts models activate parameters, compares their effects on throughput, memory cost, and serving complexity, and discusses when each architecture fits different deployment constraints. It uses Nemotron 3.5 Lightning as an example of an MoE model. ### Source excerpt How can a 30B-parameter model activate only 3B parameters per token, and still use the capacity of the larger model? Nemotron 3.5 Lightning illustrates the... ## How LLMs Handle Memory Through Context and Surrounding Applications DevFeed: [How LLMs Handle Memory Through Context and Surrounding Applications]() Original publisher: [Read original article]() Author: ByteByteGo Published: 2026-09-15T15:31:12Z Content type: article Language: en Sources: [ByteByteGo]() Topics: [Large Language Model](), [context window](), [App](), [long-context]() Tags: [article](), [context](), [context-window](), [cost](), [latency](), [llms](), [memory]() ### AI overview LLMs do not usually retain personal or persistent memory between interactions. Surrounding applications create the appearance of memory by storing messages, maintaining summaries, retrieving relevant information, and supplying it to the model. As conversations grow, this processing increases cost and latency, while context-window limits require older information to be removed, summarized, or stored elsewhere. ### Source excerpt In this article, we will learn how LLMs handle memory so that they are useful to end users in performing complex tasks that require conversation and holding context. ## How Solaris' Turnstile Influenced the Modern System Designs of Web Browsers and Language Runtimes DevFeed: [How Solaris' Turnstile Influenced the Modern System Designs of Web Browsers and Language Runtimes]() Original publisher: [Read original article]() Author: Olimpiu Pop Published: 2026-09-15T06:06:00Z Content type: article Language: en Sources: [InfoQ]() Topics: [systems](), [Programming](), [Kernel](), [observability](), [browsers](), [web browsers]() Tags: [architecture](), [architecture-design](), [architectures](), [blocking](), [concurrency](), [development](), [efficiency](), [kernel](), [locking](), [memory](), [news](), [observability](), [operating-systems](), [solaris](), [storage](), [systems](), [turnstile-system-design]() ### AI overview The article examines how Solaris innovations influenced modern systems engineering. It discusses the Slab Allocator, OpenZFS, DTrace, Zones, and turnstiles, focusing on how turnstiles reduce per-lock memory overhead while supporting priority inheritance for contended locks. ### Source excerpt Sun's Solaris influenced modern systems engineering, with key innovations like the Slab Allocator for efficient memory management, OpenZFS for advanced storage, and DTrace for observability. Its turnstile mechanism addressed issues with mutexes, promoting lean locking designs that are reflected in contemporary programming languages and web browsers, enhancing performance and memory efficiency. By Olimpiu Pop ## What It Takes to Build a Production Agent Harness DevFeed: [What It Takes to Build a Production Agent Harness]() Original publisher: [Read original article]() Author: Avi Chawla Published: 2026-09-14T19:50:37Z Content type: article Language: en Sources: [Daily Dose of Data Science]() Topics: [Agent Harness](), [LangChain](), [Artificial Intelligence](), [Persistence](), [Graphs]() Tags: [agent](), [ai](), [ai-engineer](), [engineering](), [evaluation](), [langchain](), [langgraph](), [memory](), [persistence](), [tools](), [tracing]() ### AI overview A hands-on series chapter explains how to build a production agent harness with LangChain and LangGraph. It covers model, message, prompt, and tool interactions; tool-call execution; state transitions; persistence; failure handling; tracing; evaluation; human approval; and resumable execution. ### Source excerpt A hands-on nanodegree for production agent engineering. ## MSI XpertStation WS300 Thermals: Why a 1,300W GB300 Doesn't Throttle on a Desk DevFeed: [MSI XpertStation WS300 Thermals: Why a 1,300W GB300 Doesn't Throttle on a Desk]() Original publisher: [Read original article]() Author: Brian Beeler Published: 2026-09-14T19:41:35Z Content type: article Language: en Sources: [StorageReview.com]() Topics: [Blackwell](), [cpu](), [GPU](), [Grace CPU](), [Nvidia](), [DGX Station]() Tags: [blackwell](), [consumer](), [dgx-station](), [gpu](), [grace-cpu](), [heat](), [memory](), [nvidia](), [performance](), [review](), [workstation]() ### AI overview The article explains why MSI's XpertStation WS300 can sustain a 1,300W GB300 Grace Blackwell Ultra Superchip without throttling on a desk. It attributes the thermal stability to cold plates covering the major heat-producing components, dual 360mm radiators, multiple fans, and a cooling loop rated above the system's nominal CPU and GPU load. ### Source excerpt The most common question we got about the MSI XpertStation WS300 after our review coalesces around one key theme. The GB300 Grace Blackwell Ultra Superchip is a 1,300W part that normally lives in a liquid-cooled rack, so what happens to thermals when you put it in a tower? The concern is fair: a GB300 system The post MSI XpertStation WS300 Thermals: Why a 1,300W GB300 Doesn't Throttle on a Desk appeared first on StorageReview.com. ## New hardware device can RAM into encrypted memory, expose your data DevFeed: [New hardware device can RAM into encrypted memory, expose your data]() Original publisher: [Read original article]() Author: Thomas Claburn Published: 2026-09-14T18:31:33Z Content type: news Language: en Sources: [www.theregister.com - Articles]() Topics: [Hardware](), [Security](), [Server](), [Cybersecurity]() Tags: [amd](), [confidential-computing](), [ddr5](), [ddrop](), [hardware](), [intel](), [memory](), [security](), [server]() ### AI overview A new hardware attack can compromise encrypted DDR5 memory and expose data. The attack requires physical access to the server. ### Source excerpt Attackers would need physical access to the server to pull off the DDR5 trick ## Kubernetes v1.37: Memory QoS Graduates to Beta DevFeed: [Kubernetes v1.37: Memory QoS Graduates to Beta]() Original publisher: [Read original article]() Author: Qi Wang; Sohan Kunkerkar Published: 2026-09-14T18:30:00Z Content type: article Language: en Sources: [Kubernetes Blog]() Topics: [Kubernetes](), [releases](), [configuration](), [Linux](), [Kernel]() Tags: [clusters](), [configuration](), [container](), [kernel](), [kubernetes](), [linux](), [memory](), [qos](), [releases](), [upgrade](), [v1]() ### AI overview Kubernetes v1.37 promotes Memory QoS to Beta and enables it by default on Linux nodes using cgroup v2. The article explains the updated defaults, configuration options for memory throttling and tiered memory protection, and upgrade behavior intended to preserve existing runtime behavior. ### Source excerpt Memory QoS has graduated to Beta in Kubernetes v1.37 and is now enabled by default. On Linux nodes running cgroup v2, the feature uses the memory controller to give the kernel better guidance on how to treat container memory. It was first introduced as Alpha in v1.22, and expanded in v1.36 with tiered memory reservation. This post covers what changed in v1.37, what the Beta promotion means for cluster operators, and how to configure the feature. What changed in v1.37Memory QoS is Beta and enabled by default The MemoryQoS feature gate is now Beta in v1.37. This means every v1.37 kubelet has the feature gate turned on without any configuration change. Turning on the feature by default is safe because the default kubelet configuration does not enable memory throttling or memory reservation. No memory.high, memory.min, or memory.low values are written to cgroups unless you explicitly configure them. You can opt into specific behaviors through kubelet configuration fields: Set memoryThrottlingFactor (for example, 0.9) to enable memory.high throttling on Burstable and BestEffort containers. The default is null, which means no throttling. Set memoryReservationPolicy to TieredReservation to enable tiered memory protection via memory.min and memory.low. The default is None, which means no memory reservation. Default memoryThrottlingFactor changed to null In earlier Alpha releases, memoryThrottlingFactor defaulted to 0.9, which meant enabling the feature gate caused the kubelet to set memory.high on containers. In v1.37, the default is null, so the kubelet does not set memory.high unless you configure a value. This change was made because, with the feature gate now on by default, an automatic memory.high could throttle workloads that were previously running without throttling. Making it null ensures that upgrading to v1.37 does not change runtime behavior for existing clusters. If your kubelet configuration file already contains an explicit memoryThrottlingFactor value, that ## Firefox 156 Available With Its Built-In PDF Viewer Starting Up To 45% Faster DevFeed: [Firefox 156 Available With Its Built-In PDF Viewer Starting Up To 45% Faster]() Original publisher: [Read original article]() Author: Michael Larabel Published: 2026-09-14T15:09:09Z Content type: news Language: en Sources: [Phoronix]() Topics: [Firefox](), [pdf](), [Mozilla](), [Linux](), [Nvidia]() Tags: [cpu](), [desktop-linux](), [firefox](), [history](), [linux](), [linux-benchmarking](), [linux-hardware-benchmarks](), [linux-hardware-reviews](), [linux-how-to](), [linux-performance](), [linux-server-benchmarks](), [memory](), [mozilla](), [nvidia](), [open-source-graphics](), [pdf](), [performance](), [phoronix](), [phoronix-test-suite](), [release](), [ubuntu-benchmarks](), [ubuntu-hardware](), [update](), [vulkan]() ### AI overview Firefox 156 is available with a built-in PDF viewer that starts up up to 45% faster. The release also improves memory and CPU usage when displaying large scaled-down JPEG images and includes several fixes. Vulkan Video decoding for newer NVIDIA GPUs remains limited to nightly builds. ### Source excerpt Mozilla today published their Firefox 156.0 release binaries ahead of the official Tuesday announcement... ## Announcing Redis 8.10: Compact Hash, JSONPath extensions, performance improvements, & more DevFeed: [Announcing Redis 8.10: Compact Hash, JSONPath extensions, performance improvements, & more]() Original publisher: [Read original article]() Author: Bosmat Tuvel Published: 2026-09-14T00:00:00Z Content type: release Language: en Sources: [Redis Blog]() Topics: [Redis](), [Data structures](), [Data Management](), [Open Source]() Tags: [data-management](), [improvements](), [memory](), [new-features](), [operations](), [performance](), [redis](), [streams](), [tech](), [time-series]() ### AI overview Redis 8.10 in Redis Open Source introduces compact hashes, incremental backup and restore, JSONPath syntax extensions, more flexible Stream consumption, new Set cardinality operations, atomic movement of multiple List elements, and enhanced Time Series capabilities. The release also improves memory efficiency, throughput, and operational reliability at scale. ### Source excerpt Redis 8.10 in Redis Open Source is now available, delivering improvements that make Redis more memory efficient, expressive, and easier to operate at scale. Highlights include compact hashes with up to 50% lower memory usage and 2x higher hash loadin... ## Kotlin Explicit Backing Fields: Encapsulation Tradeoffs and Downcasting Risks DevFeed: [Kotlin Explicit Backing Fields: Encapsulation Tradeoffs and Downcasting Risks]() Original publisher: [Read original article]() Author: Ehab Elwan Published: 2026-09-13T05:31:32Z Content type: opinion Language: en Sources: [ProAndroidDev - Medium]() Topics: [Kotlin](), [Code](), [Jetpack Compose]() Tags: [android](), [android-development](), [architecture](), [article](), [clean-code](), [code](), [compiler](), [kotlin](), [memory](), [software-architecture](), [stateflow]() ### AI overview This article compares Kotlin's traditional private mutable property plus public read-only wrapper with explicit backing fields. It explains that explicit backing fields can avoid an extra wrapper allocation, but because the underlying object remains mutable, an external downcast may bypass the intended read-only restriction and mutate internal state. ### Source excerpt Why eliminating the double-property boilerplate changes how we protect our architecture Image generated by AIDisclosure: This article was drafted by me and refined with the help of AI tools. If you have written Kotlin in the last few years, you are intimately familiar with the double-property boilerplate. Whether in Android ViewModels or general state holders, maintaining a private mutable property alongside a public read-only property is a chore we have all accepted in the name of strict encapsulation to prevent our internal state from being hijacked by outside classes. It makes the code significantly cleaner (and slightly more memory efficient). However, it fundamentally changes how we protect our state, moving from a physical object boundary to a simple type restriction. Let's look at the tradeoff. The Old Way: Wrapper Protection For years, the standard approach to encapsulating state has looked like this: class OldViewModel { // 1. The private mutable state private val _uiState = MutableStateFlow(UiState()) // 2. The public read-only state val uiState: StateFlow = _uiState.asStateFlow() } This is tedious to write, but it provides a strict architectural guarantee. When you call .asStateFlow(), Kotlin does not just change the type; it creates a brand new wrapper object in memory (ReadonlyStateFlow). While this physical barrier is fantastic for safety, it does mean you are incurring a minor memory allocation overhead by creating a secondary wrapper object for every exposed state. The New Way: Upcasting Explicit Backing Fields allow you to merge these two properties into one concise declaration, bypassing that extra memory allocation entirely: class NewViewModel { val uiState: StateFlow field = MutableStateFlow(UiState()) } Inside your class, the Kotlin compiler smart-casts the field so you can mutate it internally. Outside the class, the compiler restricts callers to the read-only StateFlow interface. It looks incredibly clean and saves an allocat ## EROFS Disables LZ4 Rolling Decompression Due To Data Corruption Possibility DevFeed: [EROFS Disables LZ4 Rolling Decompression Due To Data Corruption Possibility]() Original publisher: [Read original article]() Author: Michael Larabel Published: 2026-09-13T00:33:00Z Content type: news Language: en Sources: [Phoronix]() Topics: [Filesystems](), [systems](), [Optimization](), [Linux](), [data](), [Containers]() Tags: [containers](), [data](), [desktop-linux](), [embedded](), [embedded-systems](), [kernel](), [linux](), [linux-benchmarking](), [linux-hardware-benchmarks](), [linux-hardware-reviews](), [linux-how-to](), [linux-performance](), [linux-server-benchmarks](), [memory](), [open-source-graphics](), [optimization](), [phoronix](), [phoronix-test-suite](), [reduce](), [systems](), [ubuntu-benchmarks](), [ubuntu-hardware](), [x86]() ### AI overview EROFS has temporarily disabled LZ4 rolling decompression because a rare interaction with the upstream LZ4 implementation could produce corrupted data. The change prioritizes data correctness in production but increases runtime memory usage. ### Source excerpt The EROFS read-only file-system popular for embedded systems, containers, and other use-cases has resorted to temporarily disabling its LZ4 rolling decompression support due to data corruption concerns... ## Designing Reliable AI Agent Memory for Stale Facts and Policy Changes DevFeed: [Designing Reliable AI Agent Memory for Stale Facts and Policy Changes]() Original publisher: [Read original article]() Author: Raul Junco Published: 2026-09-12T12:10:58Z Content type: tutorial Language: en Sources: [System Design Classroom]() Topics: [AI Agent](), [context window](), [context]() Tags: [ai-agent](), [context](), [context-window](), [memory]() ### AI overview This article explains that AI agent memory should be treated as evidence rather than truth, especially when policies or other facts change. It recommends attaching version, scope, source, and authoritative-data checks to memory, while keeping the context window limited to the information needed for a task. ### Source excerpt Learn how to design reliable AI agent memory that handles stale facts, policy changes, scoped retrieval, conflict resolution, and safe deletion. ## How to use Google microbenchmarks for evaluating TPU performance DevFeed: [How to use Google microbenchmarks for evaluating TPU performance]() Original publisher: [Read original article]() Author: Junjie Qian; Chi Shuen Lee; Yu-Hsuan (Amy) Lin; Haixiong (Sean) Wang Published: 2026-09-12T11:04:33.891311Z Content type: tutorial Language: en Sources: [Google Developers Blog]() Topics: [Google AI](), [Inference Performance]() Tags: [architecture](), [benchmarks](), [compute](), [developers](), [google](), [guides](), [hardware](), [how-to](), [kernel](), [machine-learning](), [memory](), [mesh](), [metrics](), [model](), [open](), [open-source](), [performance](), [scale](), [software](), [tpu]() ### AI overview A tutorial on using Google's TPU microbenchmark suite to measure network, compute, memory, host-transfer, and attention performance. The results can establish a Roofline baseline and guide workload-specific optimization. ### Source excerpt Google's open-source TPU microbenchmark suite provides developers with granular performance metrics across Network, Compute, HBM, Host Transfer, and Attention components to validate real-world hardware capabilities. By leveraging these benchmarks to establish a Roofline model, engineers can accurately diagnose whether their machine learning workloads are compute-, memory-, or network-bound. This empirical baseline directly guides targeted software optimizations--such as kernel tuning, mesh sharding, and rematerialization--to maximize hardware utilization for large-scale model deployments. ## JDK 27 Runtime Updates Release Notes DevFeed: [JDK 27 Runtime Updates Release Notes]() Original publisher: [Read original article]() Author: Billy Korando Published: 2026-09-12T00:00:00Z Content type: release Language: en Sources: [Inside Java]() Topics: [JDK 27](), [Release notes](), [releases](), [Java]() Tags: [env-file-security](), [gc](), [jdk-27](), [memory](), [net-conf](), [passwords](), [performance](), [release](), [release-notes]() ### AI overview This article reviews runtime updates in JDK 27, including G1GC becoming the default garbage collector in all cases, Compact Object Headers being enabled by default, and improved JFR redaction and filtering for sensitive information such as passwords and API keys. ### Source excerpt Let's review the performance updates, new runtime features, and other changes to existing features that are in the JDK 27 release! ## What Go Taught Us About Java Garbage Collection DevFeed: [What Go Taught Us About Java Garbage Collection]() Original publisher: [Read original article]() Author: Shai Almog Published: 2026-09-12T00:00:00Z Content type: opinion Language: en Sources: [CodeName One]() Topics: [Go Language](), [Java](), [benchmarking]() Tags: [comparison](), [garbage-collection](), [go](), [java](), [latency](), [memory](), [performance]() ### AI overview A comparison between ParparVM and Go prompts an investigation into Java garbage-collection pacing, allocation thresholds, and memory use. Lowering ParparVM's allocation threshold reduced measured process resident memory from 98 MB to 38 MB, while throughput and p99 latency changed little in the reported test. ### Source excerpt A Go performance comparison led us from stack allocation to GC pacing, parallel marking, and image caches. ParparVM can explore those choices while keeping ordinary Java APIs. ## The Architecture for Serving 100 Fine-Tuned Models on One GPU DevFeed: [The Architecture for Serving 100 Fine-Tuned Models on One GPU]() Original publisher: [Read original article]() Author: Avi Chawla Published: 2026-09-11T21:25:15Z Content type: tutorial Language: en Sources: [Daily Dose of Data Science]() Topics: [GPU](), [vllm](), [Fine-tuning](), [lora](), [Serverless](), [Deployment](), [Code]() Tags: [code](), [deployment](), [fine-tuning](), [gpu](), [lora](), [memory](), [models](), [production](), [serverless](), [vllm](), [workers]() ### AI overview This tutorial compares architectures for serving 100 fine-tuned 7B model variants on GPUs. It explains how separate merged models increase storage, GPU memory use, scaling pools, cold starts, and idle capacity, while a shared base model with LoRA adapters enables adapter reuse through vLLM. The article plans to test merged, unmerged startup-loaded, request-time adapter loading, and hosted-per-tenant deployments on Runpod Serverless. ### Source excerpt ...explained with code. ## CachyOS vs. Windows 11 vs. Ubuntu 26.04 LTS On Intel Wildcat Lake + 8GB RAM DevFeed: [CachyOS vs. Windows 11 vs. Ubuntu 26.04 LTS On Intel Wildcat Lake + 8GB RAM]() Original publisher: [Read original article]() Author: Michael Larabel Published: 2026-09-11T15:08:55Z Content type: comparison Language: en Sources: [Phoronix]() Topics: [benchmarking](), [Linux](), [Ubuntu](), [Windows](), [cpu](), [Hardware](), [intel](), [Linux performance]() Tags: [benchmarking](), [desktop-linux](), [hardware](), [intel](), [intel-core](), [laptop](), [linux](), [linux-benchmarking](), [linux-hardware-benchmarks](), [linux-hardware-reviews](), [linux-how-to](), [linux-performance](), [linux-server-benchmarks](), [lts](), [memory](), [microsoft](), [nvme](), [open-source-graphics](), [performance](), [phoronix](), [phoronix-test-suite](), [release](), [review](), [ssd](), [testing](), [ubuntu](), [ubuntu-benchmarks](), [ubuntu-hardware](), [windows](), [windows-11]() ### AI overview This comparison measures Ubuntu 26.04 LTS, CachyOS, and Windows 11 performance on a $449 CHUWI UniBook laptop with an Intel Core 3 Wildcat Lake processor, 8GB of memory, and a 256GB NVMe SSD across various workloads. ### Source excerpt Often times when testing different Linux distributions or comparing Windows vs. Linux it's on leading flagship desktop or server hardware, but today we are looking at the Ubuntu vs. CachyOS vs. Windows performance at the opposite end of the spectrum. With the new CHUWI UniBook $449 laptop powered by Intel Core 3 Wildcat Lake and with 8GB of system memory, here is a look at how those three operating systems compare across a variety of workloads. ## What Is an Agent Harness? The Architecture Behind Claude Code, DeepSeek Harness, and Hermes Agent DevFeed: [What Is an Agent Harness? The Architecture Behind Claude Code, DeepSeek Harness, and Hermes Agent]() Original publisher: [Read original article]() Author: Rudrendu Paul Published: 2026-09-11T15:07:18Z Content type: tutorial Language: en Sources: [freeCodeCamp Programming Tutorials: Python, JavaScript, Git & More]() Topics: [AI Bots](), [AI-assisted coding]() Tags: [agent](), [ai-agents](), [architecture](), [artificial-intelligence](), [claude-code](), [deepseek](), [developer-tools](), [llm](), [machine-learning](), [mcp](), [memory](), [python](), [sandbox](), [software-engineering](), [tools]() ### AI overview An explainer and hands-on guide to agent harnesses: the runtime infrastructure around an LLM that manages model calls, tool execution, memory, and filesystem sandboxing. It compares popular harnesses and introduces a small Python implementation. ### Source excerpt On August 13, 2026, DeepSeek published a GitHub repository called deepseek-harness. Within two days, it had passed 95,386 stars and 8,826 forks (a vanity metric on its own, but a spike this fast signa ## Evolving Pinterest's Embedding Retrieval Platform DevFeed: [Evolving Pinterest's Embedding Retrieval Platform]() Original publisher: [Read original article]() Author: Pinterest Engineering Published: 2026-09-11T15:01:03Z Content type: article Language: en Sources: [Pinterest Engineering Blog - Medium]() Topics: [Architecture & Design](), [Algorithms](), [Scalability](), [Latency](), [IO]() Tags: [ann](), [cost](), [embedding](), [embeddings](), [engineering](), [infrastructure](), [latency](), [memory](), [models](), [pinterest](), [platform](), [quantization](), [real-time](), [retrieval](), [scalability](), [search]() ### AI overview Pinterest describes evolving its Manas embedding-retrieval platform to address the cost, scale, and flexibility challenges of serving billions of embeddings. The excerpt covers ANN search, vector quantization, and SSD-based serving. ### Source excerpt Authors: Bowen Zhou | Staff Software Engineer; Shan Gao | Senior Software Engineer; Jingwen Hu | Software Engineer II; Wenjiang Chu | Staff Software Engineer The Billion-Embedding Challenge At Pinterest, the "signal" is our lifeblood. Whether it's a home decor enthusiast finding the perfect rug or a fashion seeker discovering a new aesthetic, our discovery engine relies on understanding deep semantic relationships to help our users find inspirations. Over the last few years, the explosive growth of embedding-based retrieval has fundamentally transformed how we surface these signals -- and at the heart of that transformation is Manas, Pinterest's in-house distributed search platform. Embedding Retrieval is one of the core capabilities of Manas, supporting multiple approximate nearest neighbor search algorithms, hybrid queries with both token and embedding clauses, as well as real-time updates to ensure fresh contents become searchable within seconds. Deployed on over 80 clusters and serving billions of embeddings, Manas embedding retrieval powers all major product surfaces at Pinterest including Home Feed, Search, Related Pins, Ads, and Notifications. However, as our corpus scales toward tens of billions of embeddings and our models capture increasingly complex interactions, we face mounting challenges around cost efficiency, scalability, and flexibility. On the infrastructure side, traditional ANN algorithms like HNSW are notoriously memory-hungry -- they require the entire index to reside in RAM to maintain low query latency, making cost grow linearly with corpus size. On the modeling side, the classic two-tower retrieval paradigm is too restrictive: it reduces each candidate to a single embedding and scores relevance through a simple dot product, leaving little room to express richer, context-dependent notions of similarity. To tackle these challenges, our team has been evolving Manas's embedding retrieval stack across three fronts: Quantization. We reduce the memor ## Microsoft annoints Rust as a 'Tier 1' internal language DevFeed: [Microsoft annoints Rust as a 'Tier 1' internal language]() Original publisher: [Read original article]() Author: Joab Jackson Published: 2026-09-11T06:26:00Z Content type: news Language: en Sources: [www.theregister.com - Articles]() Topics: [Microsoft]() Tags: [bugs](), [devops](), [memory](), [microsoft](), [programming-language](), [rust](), [windows]() ### AI overview Microsoft has designated Rust a Tier 1 internal language, with the stated aim of helping developers address memory bugs in Windows. ### Source excerpt Redmond's coders get the tool to help them scrub memory bugs off Windows [Next page]()