[et_pb_section fb_built=”1″ _builder_version=”4.16″ _module_preset=”default” custom_margin=”0px||||false|false” custom_padding=”0px||||false|false” locked=”off” global_colors_info=”{}”][et_pb_row _builder_version=”4.16″ _module_preset=”default” width=”100%” max_width=”100%” custom_margin=”||||false|false” custom_padding=”0px||||false|false” global_colors_info=”{}”][et_pb_column type=”4_4″ _builder_version=”4.16″ _module_preset=”default” global_colors_info=”{}”][et_pb_text _builder_version=”4.27.6″ _module_preset=”default” header_2_line_height=”1.7em” header_3_line_height=”1.7em” custom_margin=”||-10px||false|false” custom_padding=”||0px||false|false” locked=”off” global_colors_info=”{}”]<\/p>\n
\nDiscover how Phison and MediaTek are rethinking memory architecture to unlock next-generation AI on smartphones.<\/p>\n<\/blockquote>\n
<\/p>\n
Smartphones are becoming a primary platform for AI. What began with basic on-device features has quickly evolved into support for large language models, multimodal applications, and always-on experiences that run closer to the user.<\/p>\n
Running AI locally improves responsiveness, reduces reliance on cloud infrastructure, and keeps sensitive data on the device. It also enables real-time, persistent interactions that are difficult to deliver through cloud-based models alone.<\/p>\n
<\/p>\n
Mobile AI is a defining platform shift<\/h3>\n
AI is moving to where data is created, and that increasingly means the smartphone. Devices are no longer just endpoints for AI output. They are becoming environments where models run and respond in real time.<\/p>\n
This shift is driven by three pressures: the need for low latency, greater control over data, and the rising cost of cloud-based inference at scale. Local AI addresses all three by delivering faster performance, keeping data on-device, and reducing dependence on external services.<\/p>\n
The result is a new class of always-on, context-aware experiences that operate continuously in the background. The opportunity is clear, but there is still a gap between what mobile AI promises and what today\u2019s hardware can consistently deliver.<\/p>\n
\u00a0<\/h3>\n
The constraints holding mobile generative AI back<\/h3>\n
Generative AI is inherently memory-intensive. Running AI models requires significant resources to store parameters, manage tokens, and maintain context during inference. On smartphones, those requirements quickly come up against real-world limitations.<\/p>\n
Today\u2019s most common approaches to AI training and inference, whether on a smartphone or in the cloud, present challenges on multiple fronts:<\/p>\n
\n
- \n
\n
- \n
\n
- Cloud-based AI introduces concerns around privacy, security, and ongoing token costs<\/li>\n
- Dependence on connectivity creates latency and availability issues<\/li>\n
- On-device approaches struggle with limited memory capacity and the high cost of scaling DRAM<\/li>\n
- Slow response times can degrade the user experience<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n
These constraints create a difficult tradeoff. Users can either rely on the cloud and accept its limitations or attempt on-device AI and run into performance and cost barriers.<\/p>\n
To move forward, the industry needs a different approach to how memory is used in mobile systems.<\/p>\n
\u00a0<\/h3>\n
A new approach to mobile AI architecture<\/h3>\n
Recently, Phison and MediaTek partnered to address this challenge with a fundamentally different way of thinking about memory.<\/p>\n
The joint solution combines the MediaTek Dimensity 9500 SoC with Phison\u2019s Pascari aiDAPTIV\u2122 solution, introducing a new AI inference architecture for smartphones that extends beyond DRAM limitations.<\/p>\n
At a high level, the approach is simple but powerful, as it:<\/p>\n
\n
- \n
\n
- \n
\n
- Extends the memory hierarchy by incorporating NAND flash alongside DRAM<\/li>\n
- Uses intelligent middleware to dynamically manage data across memory tiers<\/li>\n
- Treats memory and storage as a unified, coordinated resource<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n
Instead of relying solely on DRAM, aiDAPTIV leverages NAND flash as an extension of working memory, significantly expanding the available memory pool for AI workloads.<\/p>\n
The key enabler is the aiDAPTIV Memory Management Middleware, which acts as a coordination layer between the SoC, DRAM, and UFS storage. It dynamically streams model data where it\u2019s needed, effectively breaking the boundary between memory and storage.<\/p>\n
This creates a hybrid architecture where frequently accessed AI data is cached intelligently, storage is partitioned into dedicated regions for system data and AI workloads, and data can be reused and offloaded dynamically to optimize performance.<\/p>\n
In practical terms, this means your smartphone can handle larger models, longer contexts, and more complex inference tasks without requiring a dramatic increase in DRAM.<\/p>\n
\u00a0<\/h3>\n
Turn memory from a constraint into an advantage with Phison and MediaTek<\/h3>\n
This architectural shift delivers measurable benefits that directly address the core challenges of mobile AI.<\/p>\n
Reduced DRAM requirements<\/strong>
Typical mobile AI deployments may require 16 GB or more of DRAM to support advanced models or use cases, such as when leveraging mixture of experts (MoE). With dynamic model and MoE offloading and intelligent memory management, the aiDAPTIV approach can reduce those requirements to around 12 GB while maintaining performance.<\/p>\nLower system cost and improved efficiency<\/strong>
By leveraging the cost advantages of NAND flash, the Phison-MediaTek solution reduces the need for expensive DRAM scaling. This enables more cost-effective device designs without sacrificing AI capability.<\/p>\nSupport for larger models and longer context windows<\/strong>
The expanded memory pool allows your smartphone to handle more complex models and longer sequences, unlocking richer and more capable AI experiences.<\/p>\nImproved privacy and autonomy<\/strong>
Running inference locally reduces dependence on cloud infrastructure, helping protect sensitive data and enabling AI functionality even without connectivity.<\/p>\n<\/p>\n
Building the foundation for next-generation mobile AI<\/h3>\n
Part of what makes this collaboration so significant is the performance gains, but it\u2019s also about the shift in how smartphone systems are designed to support AI.<\/p>\n
By unifying memory and storage into a coordinated architecture, Phison and MediaTek are enabling a new class of smartphones that can:<\/p>\n
\n
- \n
\n
- \n
\n
- Run advanced AI models locally<\/li>\n
- Deliver faster, more responsive user experiences<\/li>\n
- Balance performance, cost, and power efficiency<\/li>\n
- Scale AI capabilities without scaling hardware complexity<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n<\/ul>\n
This is a foundational step toward truly autonomous, always-on AI at the edge.<\/p>\n
\u00a0<\/h3>\n
Looking ahead<\/h3>\n
As mobile AI continues to evolve, memory will remain one of the most important factors shaping what is possible. Solutions that rethink how memory is structured and utilized will define the next wave of innovation.<\/p>\n
The collaboration between Phison and MediaTek represents a clear direction forward. By transforming memory from a bottleneck into a scalable resource, it opens the door to more capable, efficient, and accessible AI experiences on the smartphone.<\/p>\n
[\/et_pb_text][\/et_pb_column][\/et_pb_row][et_pb_row disabled_on=”on|on|on” _builder_version=”4.16″ _module_preset=”default” width=”100%” max_width=”100%” custom_margin=”||||false|false” custom_padding=”0px||||false|false” disabled=”on” saved_tabs=”all” locked=”off” global_colors_info=”{}”][et_pb_column type=”4_4″ _builder_version=”4.16″ _module_preset=”default” global_colors_info=”{}”][et_pb_text _builder_version=”4.27.4″ _module_preset=”default” global_colors_info=”{}”]<\/p>\n
Frequently Asked Questions (FAQ) :<\/strong><\/h3>\n
[\/et_pb_text][et_pb_toggle title=”Why are AI workloads increasing pressure on GPU and DRAM supply?” _builder_version=”4.27.6″ _module_preset=”default” global_colors_info=”{}”]Modern AI models require significantly more memory for larger context windows, inference\u00a0<\/span>workloads<\/span>\u00a0and fine-tuning tasks. As\u00a0<\/span>hyperscalers<\/span> and enterprises rapidly expand AI deployments, demand for GPUs, DRAM and NAND has outpaced manufacturing capacity, creating higher costs, longer lead times and supply uncertainty across the industry.<\/span><\/span>
\n[\/et_pb_toggle][et_pb_toggle title=”What is the biggest bottleneck in enterprise AI infrastructure today?” _builder_version=”4.27.6″ _module_preset=”default” global_colors_info=”{}”]For many organizations, the biggest bottleneck is not raw compute power but inefficient data movement between storage, system\u00a0<\/span>memory<\/span>\u00a0and GPUs. When data pipelines cannot keep up with workload demands, GPUs\u00a0<\/span>remain<\/span> underutilized, reducing performance efficiency and increasing operational costs.<\/span><\/span>
\n[\/et_pb_toggle][et_pb_toggle title=”How does KV-cache impact AI inference performance?” _builder_version=”4.27.6″ _module_preset=”default” global_colors_info=”{}”]KV-cache stores token context during inference so large language models can\u00a0<\/span>maintain<\/span>\u00a0conversation continuity without repeatedly recalculating prior tokens. As context windows grow, KV-cache consumes significant GPU memory, and inefficient cache handling can increase\u00a0<\/span>recomputation<\/span>, latency and power consumption.<\/span><\/span>
\n[\/et_pb_toggle][et_pb_toggle title=”Why are Mixture-of-Experts (MoE) models memory intensive?” _builder_version=”4.27.6″ _module_preset=”default” global_colors_info=”{}”]MoE<\/span>\u00a0models rely on multiple specialized expert models that traditionally remain loaded in DRAM for fast access. As the number of experts increases, memory requirements\u00a0<\/span>rise substantially, making<\/span> infrastructure scaling more expensive and difficult for enterprise AI environments.<\/span><\/span>
\n[\/et_pb_toggle][et_pb_toggle title=”Can AI performance improve without adding more GPUs?” _builder_version=”4.27.6″ _module_preset=”default” global_colors_info=”{}”]Yes. Many AI workloads can achieve higher performance through better memory orchestration and optimized data flow rather than simply adding more GPUs. Improving GPU\u00a0<\/span>utilization<\/span>, reducing\u00a0<\/span>recomputation<\/span>\u00a0and streamlining memory access often delivers more efficient scaling at lower cost.<\/span>
\n[\/et_pb_toggle][et_pb_toggle title=”What is Phison\u2019s aiDAPTIV technology?” _builder_version=”4.27.6″ _module_preset=”default” global_colors_info=”{}”]Phison\u2019s\u00a0<\/span>aiDAPTIV<\/span>\u00a0is a controller-level AI memory orchestration platform designed to\u00a0<\/span>optimize<\/span>\u00a0how data moves between GPU memory,\u00a0<\/span>DRAM<\/span>\u00a0and high-performance flash storage. It extends effective memory capacity while improving GPU\u00a0<\/span>utilization<\/span> and reducing infrastructure inefficiencies.<\/span><\/span>
\n[\/et_pb_toggle][et_pb_toggle title=”How does aiDAPTIV reduce DRAM requirements for MoE models?” _builder_version=”4.27.6″ _module_preset=”default” global_colors_info=”{}”]aiDAPTIV<\/span>\u00a0stores less\u00a0<\/span>frequently<\/span>\u00a0used\u00a0<\/span>MoE<\/span>\u00a0experts on high-performance SSDs instead of keeping every expert permanently loaded in DRAM. Frequently accessed experts\u00a0<\/span>remain<\/span> in memory while inactive experts are retrieved with low latency only when needed, significantly lowering DRAM requirements.<\/span><\/span>
\n[\/et_pb_toggle][et_pb_toggle title=”How does aiDAPTIV improve KV-cache efficiency?” _builder_version=”4.27.6″ _module_preset=”default” global_colors_info=”{}”]aiDAPTIV<\/span>\u00a0stores evicted KV-cache tokens in flash storage instead of discarding them entirely. This allows previously used context to be retrieved quickly without forcing full\u00a0<\/span>recomputation<\/span>\u00a0on the GPU, improving latency, Time\u00a0<\/span>To<\/span> First Token performance and overall GPU efficiency.<\/span><\/span>
\n[\/et_pb_toggle][et_pb_toggle title=”What benefits does aiDAPTIV provide for enterprise AI infrastructure?” _builder_version=”4.27.6″ _module_preset=”default” global_colors_info=”{}”]aiDAPTIV<\/span>\u00a0helps enterprises improve GPU\u00a0<\/span>utilization<\/span>, reduce dependence on scarce DRAM resources, lower\u00a0<\/span>recomputation<\/span> overhead and improve inference efficiency. This enables organizations to scale AI workloads more efficiently while controlling infrastructure costs and power consumption.<\/span><\/span>
\n[\/et_pb_toggle][et_pb_toggle title=”Why is aiDAPTIV different from traditional AI scaling approaches?” _builder_version=”4.27.6″ _module_preset=”default” global_colors_info=”{}”]Traditional AI scaling often depends on\u00a0<\/span>purchasing<\/span>\u00a0<\/span>additional<\/span>\u00a0GPUs or increasing DRAM capacity.\u00a0<\/span>aiDAPTIV<\/span> instead focuses on intelligent data orchestration and tiered memory management, enabling existing hardware to deliver higher AI performance without excessive infrastructure expansion.<\/span><\/span>
\n[\/et_pb_toggle][\/et_pb_column][\/et_pb_row][\/et_pb_section]<\/p>\n","protected":false},"excerpt":{"rendered":"Discover how Phison and MediaTek are rethinking memory architecture to unlock next-generation AI on smartphones. Smartphones are becoming a primary platform for AI. What began with basic on-device features has quickly evolved into support for large language models, multimodal applications, and always-on experiences that run closer to the user. Running AI locally improves responsiveness, […]<\/p>\n","protected":false},"author":10,"featured_media":89221,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"_et_pb_use_builder":"on","_et_pb_old_content":"","_et_gb_content_width":"","inline_featured_image":false,"footnotes":""},"categories":[120,23],"tags":[22],"class_list":["post-89216","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-ai","category-all-posts","tag-long-content"],"acf":[],"_links":{"self":[{"href":"https:\/\/phisonblog.com\/ko\/wp-json\/wp\/v2\/posts\/89216","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/phisonblog.com\/ko\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/phisonblog.com\/ko\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/phisonblog.com\/ko\/wp-json\/wp\/v2\/users\/10"}],"replies":[{"embeddable":true,"href":"https:\/\/phisonblog.com\/ko\/wp-json\/wp\/v2\/comments?post=89216"}],"version-history":[{"count":10,"href":"https:\/\/phisonblog.com\/ko\/wp-json\/wp\/v2\/posts\/89216\/revisions"}],"predecessor-version":[{"id":89431,"href":"https:\/\/phisonblog.com\/ko\/wp-json\/wp\/v2\/posts\/89216\/revisions\/89431"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/phisonblog.com\/ko\/wp-json\/wp\/v2\/media\/89221"}],"wp:attachment":[{"href":"https:\/\/phisonblog.com\/ko\/wp-json\/wp\/v2\/media?parent=89216"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/phisonblog.com\/ko\/wp-json\/wp\/v2\/categories?post=89216"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/phisonblog.com\/ko\/wp-json\/wp\/v2\/tags?post=89216"}],"curies":[{"name":"\uc6cc\ub4dc\ud504\ub808\uc2a4","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}