# IBM Research Published articles for IBM Research. This is one page of public article previews, not the complete archive. Follow Next page to continue. Summaries are not the original full articles. ## Chip Industry Technical Paper Roundup: Nov. 3 DevFeed: [Chip Industry Technical Paper Roundup: Nov. 3]() Original publisher: [Read original article]() Author: Linda Christensen Published: 2026-11-03T08:01:14Z Content type: news Language: en Sources: [Semiconductor Engineering]() Topics: [Chip design]() Tags: [2-5d](), [3d](), [ai](), [analog](), [analog-placement](), [ase](), [chip](), [detection](), [embodied-emissions](), [emissions](), [eurecom](), [flip-chip](), [gpu](), [ibm-research](), [ibm-research-europe](), [imec](), [industry](), [integration](), [interconnects](), [ku-leuven](), [materials-center-leoben](), [memory](), [neural](), [neural-network](), [paper](), [pcm](), [pva-tepla](), [research](), [research-lphp](), [semiconductor-research](), [technical](), [ut-austin](), [vrije-universiteit-brussel]() ### AI overview The roundup lists recent semiconductor research on meter-scale quantum interconnects; embodied emissions in 2.5D/3D GPU integration; secure memory-dependent prefetching; routing-aware analog placement; AI-enhanced flip-chip defect detection; and superposed phase-change memory for neural-network vector storage. The excerpt names these topics but gives no methods, results, or comparisons. ### Source excerpt Meter-scale quantum interconnects; 2.5D/3D GPU integration and embodied emissions; secure memory-dependent prefetching; routing-aware analog placement; AI-enhanced flip-chip defect detection; superposed PCM for neural-network vector storage. The post Chip Industry Technical Paper Roundup: Nov. 3 appeared first on Semiconductor Engineering. ## Analog Superposition Increases Effective PCM Storage Density (IBM Research Europe) DevFeed: [Analog Superposition Increases Effective PCM Storage Density (IBM Research Europe)]() Original publisher: [Read original article]() Author: Technical Paper Link Published: 2026-10-27T20:37:31Z Content type: article Language: en Sources: [Semiconductor Engineering]() Topics: [ibm](), [AI Architecture]() Tags: [ai](), [ai-accelerators](), [ai-hardware](), [ai-ml-dl](), [analog](), [analog-computing](), [analog-memory](), [architecture](), [architectures](), [embedding-tables](), [europe](), [ibm](), [ibm-research](), [in-memory-computing](), [inference](), [memory](), [memory-density](), [networks](), [neural-networks](), [non-volatile-memory](), [paper](), [pcm](), [phase-change-memory](), [precision](), [prototype](), [published](), [storage](), [superposition](), [technical-papers]() ### AI overview IBM Research Europe researchers demonstrated a phase-change memory architecture that stores multiple AI vectors in superposition, increasing effective storage density with a tunable accuracy tradeoff. Tests on prototype PCM hardware reportedly achieved inference accuracy close to full-precision software. The supplied excerpt gives no quantitative results or implementation details. ### Source excerpt Researchers at IBM Research Europe published "Superposed analog memory for vector storage in neural networks." Summary: Researchers demonstrated an analog phase-change memory architecture that stores multiple AI vectors in superposition, increasing effective storage density with a tunable accuracy tradeoff. Tests on prototype PCM hardware showed inference accuracy close to full-precision software. Find the technical paper here.... " read more The post Analog Superposition Increases Effective PCM Storage Density (IBM Research Europe) appeared first on Semiconductor Engineering. ## IBM-backed llm-d framework adds controls for running AI models in-house DevFeed: [IBM-backed llm-d framework adds controls for running AI models in-house]() Original publisher: [Read original article]() Author: Paul Gillin Published: 2026-10-20T12:00:23Z Content type: release Language: en Sources: [SiliconANGLE]() Topics: [Self Hosted AI Tools](), [LLM optimization](), [Large Language Model](), [vllm](), [llm-reasoning](), [sglang](), [Local AI]() Tags: [accelerators](), [advanced-micro-devices](), [agentic](), [agents](), [ai](), [ai-agents](), [ai-inference](), [ai-models](), [artificial-intelligence](), [cached-context](), [distributed](), [distributed-ai-inference-framework](), [enterprises](), [ibm](), [ibm-corp](), [ibm-research](), [inference](), [inference-costs](), [infrastructure](), [large](), [llm-d](), [model-routing](), [multi-tenant-quality-of-service](), [news](), [on-premises-ai](), [open-source](), [pytorch](), [pytorch-conference](), [red-hat](), [release](), [the-latest]() ### AI overview llm-d 1.0 adds Kubernetes-based operational controls for running distributed model inference on an organization's own infrastructure: multi-tenant quality of service, request flow control and prioritization, and upgrades without downtime. For agent workloads, it routes requests toward replicas with reusable cached context and can move cached information between GPU memory, CPU memory, and storage. The release integrates these capabilities with prompt/response separation, batch optimization, and distribution of mixture-of-experts models across accelerators. It works with inference engines such as vLLM and SGLang. IBM reported 97% cache reuse in an agentic coding test; IBM and Red Hat separately said a demonstration ran 3,000 coding agents at five to 10 times lower cost than a hosted service using the same open model. These are company-reported results. ### Source excerpt IBM Corp. and several other contributors to the open-source llm-d project are releasing version 1.0 of their distributed artificial intelligence inference framework, adding operational controls intended to help enterprises run large models reliably on their own infrastructure. The release, announced ahead of the PyTorch Conference North America in San Jose today and tomorrow, focuses on [...] The post IBM-backed llm-d framework adds controls for running AI models in-house appeared first on SiliconANGLE. ## New course teaches how AI models run on modern accelerators DevFeed: [New course teaches how AI models run on modern accelerators]() Original publisher: [Read original article]() Author: Kaoutar El Maghraoui Published: 2026-10-19T12:00:00Z Content type: article Language: en Sources: [IBM Research]() Topics: [Learning](), [NVIDIA Research](), [coreboot]() Tags: [accelerators](), [ai](), [ai-hardware](), [ai-models](), [ai-systems](), [algorithms](), [c](), [c-plus-plus](), [cerebras](), [columbia-university](), [computer-science](), [generative-ai](), [ibm-research](), [labs](), [research](), [specialization](), [technical-note]() ### AI overview IBM Research describes a six-course Coursera specialization that teaches engineers how AI models execute across five accelerator architectures: GPUs, Google TPUs, Cerebras WSE, Groq LPU, and IBM Spyre. The courses cover execution models, compilation, memory systems, debugging, and performance modeling, with hands-on labs using PyTorch. Core labs run on Google Colab; optional practice uses Spyre hardware. ### Source excerpt 'Dataflow Architectures for AI' is a 6-course specialization from IBM Research that compares leading accelerator architectures through hands-on labs. ## Chip Industry Week in Review: Fab expansion, AI hardware, and quantum advances DevFeed: [Chip Industry Week in Review: Fab expansion, AI hardware, and quantum advances]() Original publisher: [Read original article]() Author: The SE Staff Published: 2026-10-16T07:01:05Z Content type: news Language: en Sources: [Semiconductor Engineering]() Topics: [Chip design](), [High-Performance Computing](), [AI Infrastructure](), [Post-quantum cryptography](), [tsmc](), [Detection engineering]() Tags: [applied-materials](), [ase](), [asml](), [ayar-labs](), [business](), [business-news](), [chip](), [chipagents](), [cisco](), [compute](), [computing-news](), [cxmt](), [d-wave](), [faraday-technology](), [formfactor](), [fujifilm](), [global-unichip-corp](), [globalfoundries](), [ibm](), [ibm-research](), [ic-link-by-imec](), [idc](), [infineon](), [intel](), [kaist](), [keysight](), [langdao-silex](), [lattice](), [marvell](), [microchip](), [microsoft](), [mueon](), [news](), [nikon](), [nvidia](), [quantum](), [semiconductor](), [startups](), [week-in-review]() ### AI overview This semiconductor industry roundup covers fab expansion and investment, advances in memory and EUV, chip design tools and AI infrastructure, security developments, and quantum computing. It reports that AI demand is driving equipment spending and startup funding, while China-based companies are advancing memory and lithography prototypes. The article also describes new chip design automation research and security initiatives, alongside continuing constraints in manufacturing capacity and supply chains. ### Source excerpt Fab expansion accelerates; China advances memory and EUV; Teradyne's test deal; SEMICON West and OCP highlights; 3D SiP; TSMC and ASML signals; 300mm fab spending; optical and 3D-compute startups raise fresh funding; quantum HW advances. The post Chip Industry Week In Review appeared first on Semiconductor Engineering. ## How llm-d makes the most of the hardware you already have DevFeed: [How llm-d makes the most of the hardware you already have]() Original publisher: [Read original article]() Author: Peter Hess Published: 2026-09-08T12:00:00Z Content type: article Language: en Sources: [IBM Research]() Topics: [LLM optimization](), [Large Language Model](), [Self Hosted AI Tools](), [cloud-infrastructure](), [cloud cost management]() Tags: [agentic](), [ai](), [ai-for-code](), [coding-agents](), [cost](), [generative-ai](), [glm-5-2](), [gpus](), [hybrid-cloud](), [hybrid-cloud-platform](), [ibm-research](), [llm-d](), [news](), [open-source](), [red-hat](), [scaling-ai]() ### AI overview In a benchmark deployment, llm-d served the approximately 753-billion-parameter GLM-5.2 open-weight model (about 39 billion active parameters) on 544 NVIDIA H100 GPUs, handling up to 3,000 concurrent coding agents with zero preemptions. On workloads with hundreds of agent sessions and high context reuse, it reached a peak of more than 6.6 million output tokens per minute. The excerpt reports self-hosting costs 5-10 times less per token than equivalent commercial API pricing at current cloud rental rates, with the largest savings on input-heavy agent workloads. These are results from a demonstration, not evidence of performance across all deployments. ### Source excerpt IBM Research and Red Hat deployed a 753B open model on H100 GPUs, serving thousands of concurrent coding agents at 5-10x lower cost than commercial APIs.