{"id":13975,"date":"2026-11-17T08:58:24","date_gmt":"2026-11-16T23:58:24","guid":{"rendered":"https:\/\/news.skhynix.com\/en\/?p=13975"},"modified":"2026-11-17T08:58:24","modified_gmt":"2026-11-16T23:58:24","slug":"ai-ecosystem-series-ep4-5","status":"publish","type":"post","link":"https:\/\/news.skhynix.com\/en\/ai-ecosystem-series-ep4-5\/","title":{"rendered":"[AI Ecosystem] Completing AI: Physical intelligence and the role of memory"},"content":{"rendered":"<div class=\"post-intro\">\n<p>As AI workloads evolve toward inference and agentic AI, the criteria for system performance are also changing. Computing power alone is no longer enough. Where data is stored, how quickly it moves, and how efficiently it is processed now have a direct impact on the performance and efficiency of AI systems.<\/p>\n<div style=\"height: 16px; line-height: 4px;\"><\/div>\n<p>This article series explores the transformation of the AI ecosystem, where software, datacenter infrastructure, semiconductors, and memory technologies work together. It also examines why memory semiconductors have become a foundational technology for the AI era and outlines SK hynix\u2019s vision for building the next-generation AI memory ecosystem.<\/p>\n<div style=\"height: 16px; line-height: 4px;\"><\/div>\n<p>Part 5 examines why memory and system design must evolve together as AI expands beyond datacenters into the physical world, through insights from Nam Sung Kim, Endowed Chair Professor at the University of Illinois Urbana-Champaign(UIUC), who has researched low-power processor architectures and led the development of next-generation DRAM and processing-in-memory(PIM) technology.<\/p>\n<div style=\"height: 16px; line-height: 4px;\"><\/div>\n<p>\u2460 The paradigm shift in AI computing<br \/>\n\u2461 The real bottleneck: Data, not compute<br \/>\n\u2462 Redesigning infrastructure: Why architecture determines performance<br \/>\n\u2463 The semiconductor paradigm shifts toward memory<br \/>\n<strong>\u2464 Completing AI: Physical intelligence and the role of memory \u2014 Professor Nam-sung Kim, University of Illinois Urbana-Champaign(UIUC)<\/strong><\/p>\n<\/div>\n<div style=\"height: 16px; line-height: 16px;\"><\/div>\n<div style=\"text-align: center;\">\n<div style=\"display: inline-block; max-width: 748px; width: 100%; text-align: left;\">\n<div style=\"height: 2px; background: #666666; margin-bottom: 6px;\"><\/div>\n<h3 class=\"sub-title\" style=\"margin: 0; line-height: 1.4;\">Physical AI and memory-centric system-level design<\/h3>\n<div style=\"height: 2px; background: #666666; margin-top: 6px;\"><\/div>\n<\/div>\n<\/div>\n<p>In the era of physical AI, memory subsystem architecture and placing data across the memory hierarchy are becoming just as important as system architecture and computing that data. As AI expands beyond models confined to datacenters into systems that interact with the physical world in real time, such as robots, vehicles, industrial drones, AI smart glasses, and wearable devices, improving compute performance alone is no longer sufficient. Performance increasingly depends on where data is stored, how close to its source it can be processed, and how efficiently it can be delivered when needed with minimal power consumption.<\/p>\n<p>Parts 1 through 4 of this series explored how the factors determining AI performance are expanding beyond raw compute capability to include data delivery, infrastructure architecture, and memory-centric system design. Part 5 shifts the focus beyond the datacenter. As AI moves into the physical world, constraints involving power consumption, heat generation, response time, and device size are becoming more demanding. To ensure reliable AI operation within these constraints, the very starting point of system design must evolve.<\/p>\n<p>Smart glasses must rapidly interpret information within the user\u2019s field of view, while robots need to perceive their surroundings and adjust their movements in near real time. Similarly, vehicles and industrial drones must also make autonomous decisions, even when connectivity is lost or response times lag. In such situations, offloading all decision-making to a datacenter is impractical. Instead, a system architecture that processes data close to where it is generated and makes decisions locally is better suited to the task.<\/p>\n<div><video class=\"alignnone size-full\" style=\"display: block; width: 100%; height: auto;\" autoplay=\"autoplay\" loop=\"loop\" muted=\"\" width=\"1600\" height=\"910\"><span data-mce-type=\"bookmark\" style=\"display: inline-block; width: 0px; overflow: hidden; line-height: 0;\" class=\"mce_SELRES_start\">\ufeff<\/span><br \/>\n<span data-mce-type=\"bookmark\" style=\"display: inline-block; width: 0px; overflow: hidden; line-height: 0;\" class=\"mce_SELRES_start\">\ufeff<\/span><source src=\"https:\/\/d18r0a86za96sg.cloudfront.net\/wp-content\/uploads\/2026\/11\/03093332\/AI-Ecosystem-Completing-AI-Physical-intelligence-and-the-role-of-memory_01_motion_2026.webm\" type=\"video\/webm\" \/><span data-mce-type=\"bookmark\" style=\"display: inline-block; width: 0px; overflow: hidden; line-height: 0;\" class=\"mce_SELRES_end\">\ufeff<\/span><\/video><\/div>\n<div style=\"height: 16px; line-height: 16px;\"><\/div>\n<div style=\"text-align: center;\">\n<div style=\"display: inline-block; max-width: 748px; width: 100%; text-align: left;\">\n<div style=\"height: 2px; background: #666666; margin-bottom: 6px;\"><\/div>\n<h3 class=\"sub-title\" style=\"margin: 0; line-height: 1.4;\">Physical constraints facing AI beyond the datacenter<\/h3>\n<div style=\"height: 2px; background: #666666; margin-top: 6px;\"><\/div>\n<\/div>\n<\/div>\n<p>Until now, high-performance AI has largely operated within hyperscale datacenters. These facilities offer access to large-scale GPU clusters, ample board space, substantial power supplies, and robust cooling systems \u2014 providing sufficient system resources to handle demanding computational workloads and complex data movement.<\/p>\n<p>However, the environments in which physical AI and on-device AI<span style=\"color: #ff0000;\">*<\/span> operate differ significantly from datacenters. On-device AI integrates AI capabilities directly into devices close to users, such as smartphones, smart glasses, and wearables. Both physical AI and on-device AI require high performance, low latency, and high energy efficiency simultaneously.<\/p>\n<div class=\"footnote\"><span style=\"color: red;\">* <\/span>On-device AI: An approach in which AI computations are performed directly on devices such as smartphones, wearables, and smart glasses without sending data to external servers<\/div>\n<p>When AI is integrated into physical devices, however, battery capacity, heat dissipation, package size, and interconnect architecture become critical design constraints. Edge and mobile environments cannot easily accommodate large GPU clusters or expand standard memory modules as needed. Energy resources are also limited, as is the capacity to dissipate heat. As devices become smaller, their power budgets shrink, and the space available for heat dissipation becomes increasingly restricted.<\/p>\n<p>As a result, AI operating in the physical world must process data locally within the device whenever possible, or at least close to where it is generated. Robotic movements, autonomous vehicle driving decisions, and industrial control systems are highly sensitive to latency. Longer network round-trip times can delay critical decision-making, while unstable connections compromise the overall reliability of the system. As AI moves beyond the datacenter, product design must begin by considering memory capacity, data transfer paths, and latency.<\/p>\n<div style=\"height: 16px; line-height: 16px;\"><\/div>\n<div style=\"text-align: center;\">\n<div style=\"display: inline-block; max-width: 748px; width: 100%; text-align: left;\">\n<div style=\"height: 2px; background: #666666; margin-bottom: 6px;\"><\/div>\n<h3 class=\"sub-title\" style=\"margin: 0; line-height: 1.4;\">How algorithms are evolving to address memory challenges<\/h3>\n<div style=\"height: 2px; background: #666666; margin-top: 6px;\"><\/div>\n<\/div>\n<\/div>\n<p>These constraints are also influencing algorithm design. One approach is model compression, which reduces model size while minimizing performance degradation, thereby expanding the range of models that can run directly on devices. One widely used technique is quantization<span style=\"color: #ff0000;\">*<\/span>, which represents model parameters using fewer bits to reduce storage requirements and the burden of data transfer.<\/p>\n<p>For physical AI and on-device AI, however, simply reducing model size is not enough. Inference performance must also be maintained even under tight power and thermal constraints. TurboQuant<span style=\"color: #ff0000;\">*<\/span> is one example of a quantization technique designed to address these requirements, focusing on reducing memory demands during inference while preserving the quality of model responses.<\/p>\n<p>More recently, approaches have emerged that prioritize memory lookups over repetitive computation. Engram<span style=\"color: #ff0000;\">*<\/span> is one such example. In conventional large language models(LLMs), the system typically processes information or reconstructs patterns by passing through multiple neural network layers sequentially. Engram, by contrast, stores frequently used patterns and information in a \u201cconditional memory\u201d lookup table, retrieving only the information relevant to a given input. This is similar to how people make notes of frequently used information and refer back to them when needed. While this approach reduces the computational load \u2014 as the same information does not need to be recalculated each time \u2014 it increases the amount of data that must be stored and retrieved. As a result, memory capacity and data delivery capabilities become the primary constraints on processing speed and efficiency.<\/p>\n<div class=\"footnote\"><span style=\"color: red;\">* <\/span>Quantization: A technique that represents model weights, activation values, and key-value(KV) caches using fewer bits to reduce memory usage and data movement. For example, representing 16-bit data in 8-bit or 4-bit formats can reduce storage requirements and data transfer overhead<br \/>\n<span style=\"color: red;\">* <\/span>TurboQuant: A quantization technique that reduces memory usage by compressing the KV cache into lower-bit representations during LLM inference<br \/>\n<span style=\"color: red;\">* <\/span>Engram: An architecture that uses conditional memory and large-scale lookup tables to shift some computational workloads to memory lookups, rather than reconstructing information solely through computations within internal neural network calculations<\/div>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"alignnone size-full wp-image-14036\" src=\"https:\/\/d18r0a86za96sg.cloudfront.net\/wp-content\/uploads\/2026\/11\/12135047\/AI-Ecosystem-Completing-AI-Physical-intelligence-and-the-role-of-memory_02-1_graphic_2026.jpg\" alt=\"\" width=\"3333\" height=\"2417\" srcset=\"https:\/\/d18r0a86za96sg.cloudfront.net\/wp-content\/uploads\/2026\/11\/12135047\/AI-Ecosystem-Completing-AI-Physical-intelligence-and-the-role-of-memory_02-1_graphic_2026.jpg 3333w, https:\/\/d18r0a86za96sg.cloudfront.net\/wp-content\/uploads\/2026\/11\/12135047\/AI-Ecosystem-Completing-AI-Physical-intelligence-and-the-role-of-memory_02-1_graphic_2026-300x218.jpg 300w, https:\/\/d18r0a86za96sg.cloudfront.net\/wp-content\/uploads\/2026\/11\/12135047\/AI-Ecosystem-Completing-AI-Physical-intelligence-and-the-role-of-memory_02-1_graphic_2026-1024x743.jpg 1024w, https:\/\/d18r0a86za96sg.cloudfront.net\/wp-content\/uploads\/2026\/11\/12135047\/AI-Ecosystem-Completing-AI-Physical-intelligence-and-the-role-of-memory_02-1_graphic_2026-768x557.jpg 768w, https:\/\/d18r0a86za96sg.cloudfront.net\/wp-content\/uploads\/2026\/11\/12135047\/AI-Ecosystem-Completing-AI-Physical-intelligence-and-the-role-of-memory_02-1_graphic_2026-1536x1114.jpg 1536w, https:\/\/d18r0a86za96sg.cloudfront.net\/wp-content\/uploads\/2026\/11\/12135047\/AI-Ecosystem-Completing-AI-Physical-intelligence-and-the-role-of-memory_02-1_graphic_2026-2048x1485.jpg 2048w, https:\/\/d18r0a86za96sg.cloudfront.net\/wp-content\/uploads\/2026\/11\/12135047\/AI-Ecosystem-Completing-AI-Physical-intelligence-and-the-role-of-memory_02-1_graphic_2026-1200x870.jpg 1200w, https:\/\/d18r0a86za96sg.cloudfront.net\/wp-content\/uploads\/2026\/11\/12135047\/AI-Ecosystem-Completing-AI-Physical-intelligence-and-the-role-of-memory_02-1_graphic_2026-1980x1436.jpg 1980w\" sizes=\"auto, (max-width: 3333px) 100vw, 3333px\" \/><\/p>\n<p>Although TurboQuant and Engram take different approaches, both represent attempts to recalibrate the relationship between computation and memory. Software optimization alone has limitations in improving real-world performance. As AI expands into the physical world, changes in algorithms increasingly drive changes in memory architecture.<\/p>\n<p>As AI moves beyond datacenters into edge and mobile devices, the requirements for memory design are also evolving. In physical AI applications in particular, any latency in data processing can translate directly into delays in real-world decision-making and physical actions. Consequently, the scope of design expands beyond memory itself to encompass system-level integration \u2014 covering everything from the placement and interconnection of memory and logic components to securing sufficient bandwidth within a package and managing the resulting heat during operation.<\/p>\n<div style=\"height: 16px; line-height: 16px;\"><\/div>\n<div style=\"text-align: center;\">\n<div style=\"display: inline-block; max-width: 748px; width: 100%; text-align: left;\">\n<div style=\"height: 2px; background: #666666; margin-bottom: 6px;\"><\/div>\n<h3 class=\"sub-title\" style=\"margin: 0; line-height: 1.4;\">From HBM to HBF: The technologies shaping memory-centric architectures<\/h3>\n<div style=\"height: 2px; background: #666666; margin-top: 6px;\"><\/div>\n<\/div>\n<\/div>\n<p>Memory design for physical AI must address three key challenges: placing data close to compute devices, flexibly connecting the memory resources needed, and storing and supplying large volumes of data at the appropriate memory layer. Different technologies are used to address these challenges, including custom HBM<span style=\"color: #ff0000;\">*<\/span>, 3D-stacked DRAM<span style=\"color: #ff0000;\">*<\/span>, custom DRAM, Processing-in-Memory<span style=\"color: #ff0000;\">*<\/span>, Compute Express Link<span style=\"color: #ff0000;\">*<\/span>, High Bandwidth Flash<span style=\"color: #ff0000;\">*<\/span>, and Solid-State Drives(SSDs).<\/p>\n<div class=\"footnote\"><span style=\"color: red;\">* <\/span>Custom HBM: An HBM architecture tailored to specific AI workloads or system requirements by adjusting bandwidth, capacity, power characteristics, and packaging configurations<br \/>\n<span style=\"color: red;\">* <\/span>3D-stacked DRAM: A technology that vertically stacks DRAM directly on top of logic semiconductors, such as systems-on-chip(SoC), to shorten data transfer paths and increase input\/output(I\/O) connectivity. It focuses on reducing latency and power consumption while improving space utilization<br \/>\n<span style=\"color: red;\">* <\/span>Processing-in-memory(PIM): A technology that performs certain computations within or near memory<br \/>\n<span style=\"color: red;\">* <\/span>Compute Express Link(CXL): A high-speed interconnect standard that connects processors, accelerators, and memory devices to enable memory expansion and sharing<br \/>\n<span style=\"color: red;\">* <\/span>High Bandwidth Flash(HBF): A next-generation NAND-based memory technology being explored as a new memory layer between HBM and SSDs. Like HBM, HBF applies logic processing technology to a base die, aiming to provide significantly greater capacity than HBM while achieving comparable bandwidth and cost efficiency<\/div>\n<p>First, custom HBM, 3D-stacked DRAM, and custom DRAM focus on placing memory closer to compute devices while tailoring capacity, bandwidth, and power characteristics to suit specific workloads and system requirements. HBM vertically stacks multiple DRAM dies and connects them through a wide interface to deliver high bandwidth within a given footprint. 3D-stacked DRAM further reduces the distance between memory and compute devices by stacking DRAM directly on top of logic chips or systems-on-chip(SoC), increasing data transfer capacity within a limited space. Custom DRAM similarly optimizes memory characteristics for specific form factors<span style=\"color: #ff0000;\">*<\/span> and workloads.<\/p>\n<div class=\"footnote\"><span style=\"color: red;\">* <\/span>Form factor: The physical design constraints of a device, including its size, shape, internal space, power requirements, and thermal conditions<\/div>\n<p>PIM takes this approach a step further by moving certain computations closer to where data resides. As the volume of data transferred between memory and processors increases, so do energy consumption and latency. For workloads involving frequent, repetitive data access, PIM performs certain processing operations within or near memory to reduce unnecessary data movement.<\/p>\n<p>In cases where placing memory closer to computation is not enough, the connection architecture must also change. For example, systems such as Engram, which store and retrieve large amounts of information through lookup tables, or physical AI agents that maintain extensive contextual and state information may exceed the capacity limits of a device\u2019s internal memory. CXL addresses this by providing a high-speed, serial interface between processors, accelerators, and memory devices, enabling the expansion and sharing of external memory resources.<\/p>\n<p>Because not all data can reside in HBM or DRAM, large volumes of data are handled by the storage layer. Robot sensor logs, autonomous driving map data, and information accumulated over extended periods in physical environments can be stored in 3D NAND-based SSDs and retrieved into DRAM or HBM when needed. However, large datasets that are frequently accessed during inference, such as model weights, require higher bandwidth than conventional SSDs can provide. HBF targets this specific need as a new memory layer, stacking 3D NAND dies and using a wide interface to combine NAND\u2019s large storage capacity with high data transfer speeds.<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"alignnone size-full wp-image-13969\" src=\"https:\/\/d18r0a86za96sg.cloudfront.net\/wp-content\/uploads\/2026\/11\/03093212\/AI-Ecosystem-Completing-AI-Physical-intelligence-and-the-role-of-memory_03_graphic_2026.png\" alt=\"\" width=\"1600\" height=\"1160\" srcset=\"https:\/\/d18r0a86za96sg.cloudfront.net\/wp-content\/uploads\/2026\/11\/03093212\/AI-Ecosystem-Completing-AI-Physical-intelligence-and-the-role-of-memory_03_graphic_2026.png 1600w, https:\/\/d18r0a86za96sg.cloudfront.net\/wp-content\/uploads\/2026\/11\/03093212\/AI-Ecosystem-Completing-AI-Physical-intelligence-and-the-role-of-memory_03_graphic_2026-300x218.png 300w, https:\/\/d18r0a86za96sg.cloudfront.net\/wp-content\/uploads\/2026\/11\/03093212\/AI-Ecosystem-Completing-AI-Physical-intelligence-and-the-role-of-memory_03_graphic_2026-1024x742.png 1024w, https:\/\/d18r0a86za96sg.cloudfront.net\/wp-content\/uploads\/2026\/11\/03093212\/AI-Ecosystem-Completing-AI-Physical-intelligence-and-the-role-of-memory_03_graphic_2026-768x557.png 768w, https:\/\/d18r0a86za96sg.cloudfront.net\/wp-content\/uploads\/2026\/11\/03093212\/AI-Ecosystem-Completing-AI-Physical-intelligence-and-the-role-of-memory_03_graphic_2026-1536x1114.png 1536w, https:\/\/d18r0a86za96sg.cloudfront.net\/wp-content\/uploads\/2026\/11\/03093212\/AI-Ecosystem-Completing-AI-Physical-intelligence-and-the-role-of-memory_03_graphic_2026-1200x870.png 1200w\" sizes=\"auto, (max-width: 1600px) 100vw, 1600px\" \/><\/p>\n<p>Although these technologies serve different roles, they share a common goal: placing data where it is needed and reducing unnecessary data movement. HBM, 3D-stacked DRAM and PIM optimize memory layers close to computation, while CXL provides flexible connectivity for additional required memory resources. Meanwhile, SSDs and HBF store and supply larger volumes of data. Ultimately, memory design for physical AI comes down to how these different layers are combined to meet the requirements of specific workloads and systems.<\/p>\n<div style=\"height: 16px; line-height: 16px;\"><\/div>\n<div style=\"text-align: center;\">\n<div style=\"display: inline-block; max-width: 748px; width: 100%; text-align: left;\">\n<div style=\"height: 2px; background: #666666; margin-bottom: 6px;\"><\/div>\n<h3 class=\"sub-title\" style=\"margin: 0; line-height: 1.4;\">System-level AI memory solutions and the role of SK hynix<\/h3>\n<div style=\"height: 2px; background: #666666; margin-top: 6px;\"><\/div>\n<\/div>\n<\/div>\n<p>As AI expands beyond datacenters into edge devices, robots, and vehicles, it is becoming increasingly difficult to meet the needs of every workload with a single set of memory specifications. Large-scale training and inference demand high bandwidth and capacity, while on-device AI in compact form factors prioritizes energy efficiency and low latency. At the same time, as the volume of data generated from the field\u2014such as robot sensor logs, autonomous driving map data, and more \u2014 continues to grow, designing the storage layer has also emerged as an important challenge.<\/p>\n<p>Consequently, a diverse range of memory and storage technologies, from HBM to HBF, must be integrated into a unified portfolio at the AI system level. The benchmark for competitiveness in AI memory is also shifting from the performance of individual products to how effectively the entire memory hierarchy can be configured for specific workloads and form factors.<\/p>\n<p>This is the heart of SK hynix\u2019s vision for full-stack AI memory solutions. Achieving system-level performance and energy efficiency requires bringing together HBM and other DRAM products, including 3D-stacked DRAM, custom memory, SSDs, HBF, CXL, and PIM through a unified, full-stack approach to AI system design. Enabling these choices and combinations requires Memory Library capabilities that support memory configurations tailored to customers\u2019 architectures and workloads.<\/p>\n<p>Beyond merely supplying products tailored to customer requirements, SK hynix aims to become an architecture partner, co-designing memory configurations optimized for each customer\u2019s specific architecture. Future AI systems must be designed with consideration not only for computing capabilities but also the entire lifecycle of data: where it is stored, how it moves, and where it is processed. As the scope of physical AI expands, memory is becoming an integral part of system design, ensuring the timely supply of the right data while bridging the gap between compute devices and storage.<\/p>\n<p>As AI moves further into the physical world, how data is stored, transferred, and processed directly affects system responsiveness and efficiency. With memory playing an increasingly decisive role in AI system performance, we are entering an era in which the evolution of memory will determine the evolution of AI itself.<\/p>\n<p><img loading=\"lazy\" decoding=\"async\" class=\"alignnone size-full wp-image-13970\" src=\"https:\/\/d18r0a86za96sg.cloudfront.net\/wp-content\/uploads\/2026\/11\/03093228\/AI-Ecosystem-Completing-AI-Physical-intelligence-and-the-role-of-memory_04_etc_2026.png\" alt=\"\" width=\"1600\" height=\"484\" srcset=\"https:\/\/d18r0a86za96sg.cloudfront.net\/wp-content\/uploads\/2026\/11\/03093228\/AI-Ecosystem-Completing-AI-Physical-intelligence-and-the-role-of-memory_04_etc_2026.png 1600w, https:\/\/d18r0a86za96sg.cloudfront.net\/wp-content\/uploads\/2026\/11\/03093228\/AI-Ecosystem-Completing-AI-Physical-intelligence-and-the-role-of-memory_04_etc_2026-300x91.png 300w, https:\/\/d18r0a86za96sg.cloudfront.net\/wp-content\/uploads\/2026\/11\/03093228\/AI-Ecosystem-Completing-AI-Physical-intelligence-and-the-role-of-memory_04_etc_2026-1024x310.png 1024w, https:\/\/d18r0a86za96sg.cloudfront.net\/wp-content\/uploads\/2026\/11\/03093228\/AI-Ecosystem-Completing-AI-Physical-intelligence-and-the-role-of-memory_04_etc_2026-768x232.png 768w, https:\/\/d18r0a86za96sg.cloudfront.net\/wp-content\/uploads\/2026\/11\/03093228\/AI-Ecosystem-Completing-AI-Physical-intelligence-and-the-role-of-memory_04_etc_2026-1536x465.png 1536w, https:\/\/d18r0a86za96sg.cloudfront.net\/wp-content\/uploads\/2026\/11\/03093228\/AI-Ecosystem-Completing-AI-Physical-intelligence-and-the-role-of-memory_04_etc_2026-1200x363.png 1200w\" sizes=\"auto, (max-width: 1600px) 100vw, 1600px\" \/><\/p>\n<p style=\"text-align: left;\"><em><strong>Disclaimer:<\/strong> <span style=\"color: #808080;\">The opinions expressed in this article are solely those of the author and do not necessarily reflect the official position of SK hynix.<\/span><\/em><\/p>\n","protected":false},"excerpt":{"rendered":"<p>As AI workloads evolve toward inference and agentic AI, the criteria for system performance are also changing. Computing power alone is no longer enough. Where data is stored, how quickly it moves, and how efficiently it is processed now have<\/p>\n","protected":false},"author":47,"featured_media":13960,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"_migrated_source_id":0,"footnotes":"","_members_access_role":[],"_members_access_error":""},"categories":[5],"tags":[1642,173,451,28,177],"class_list":["post-13975","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-tech-and-ai","tag-ai-data","tag-ai-data-center","tag-ai-ecosystem","tag-physical-ai","tag-pim"],"acf":[],"_links":{"self":[{"href":"https:\/\/news.skhynix.com\/en\/wp-json\/wp\/v2\/posts\/13975","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/news.skhynix.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/news.skhynix.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/news.skhynix.com\/en\/wp-json\/wp\/v2\/users\/47"}],"replies":[{"embeddable":true,"href":"https:\/\/news.skhynix.com\/en\/wp-json\/wp\/v2\/comments?post=13975"}],"version-history":[{"count":11,"href":"https:\/\/news.skhynix.com\/en\/wp-json\/wp\/v2\/posts\/13975\/revisions"}],"predecessor-version":[{"id":14037,"href":"https:\/\/news.skhynix.com\/en\/wp-json\/wp\/v2\/posts\/13975\/revisions\/14037"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/news.skhynix.com\/en\/wp-json\/wp\/v2\/media\/13960"}],"wp:attachment":[{"href":"https:\/\/news.skhynix.com\/en\/wp-json\/wp\/v2\/media?parent=13975"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/news.skhynix.com\/en\/wp-json\/wp\/v2\/categories?post=13975"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/news.skhynix.com\/en\/wp-json\/wp\/v2\/tags?post=13975"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}