{"id":226290,"date":"2026-11-12T09:09:09","date_gmt":"2026-11-12T09:09:09","guid":{"rendered":"https:\/\/www.europesays.com\/korea\/226290\/"},"modified":"2026-11-12T09:09:09","modified_gmt":"2026-11-12T09:09:09","slug":"sk-hynix-unveils-memory-pool-concept-to-break-through-gpu-wall-co-designed-with-intel-biggo-finance","status":"publish","type":"post","link":"https:\/\/www.europesays.com\/korea\/226290\/","title":{"rendered":"SK Hynix Unveils &#8216;Memory Pool&#8217; Concept to Break Through GPU Wall, Co-Designed With Intel \u2014 BigGo Finance"},"content":{"rendered":"<p>South Korea&#8217;s SK Hynix has put front and center a so-called &#8220;memory pool&#8221; architecture in which multiple servers jointly use large-capacity memory, in a bid to resolve the memory bottleneck in artificial intelligence (AI) systems. The idea departs from the conventional approach of attaching memory to each individual GPU (graphics processing unit) and instead separates compute resources from storage resources.<\/p>\n<p>In a keynote presentation on the 12th at the &#8220;Intel AI Summit Seoul 2026&#8221; held at COEX in Gangnam District, Seoul, SK Hynix Vice President Joo Young-pyo said, &#8220;As we move into the era of agentic AI, we need to think about whether pressure on memory capacity and performance will ease or intensify,&#8221; adding, &#8220;We are on the side that believes it will intensify.&#8221; He continued, &#8220;We are discussing with customers and partners from the system design stage, and we are carrying out this work closely with Intel as well.&#8221;<\/p>\n<p>The problem he identified is physical space. A GPU has a compute core in the center, with interfaces on either side and memory attached above and below, leaving no room to stack more on the two-dimensional plane. That means it is difficult to increase capacity and bandwidth simultaneously. SK Hynix&#8217;s summary of the three major bottlenecks in AI systems is insufficient memory capacity, delayed data access, and the complexity of data management.<\/p>\n<p>When memory is scattered across individual GPUs, delays arise in the process of exchanging data, and when storage space runs short, previously computed data must be discarded and recalculated. &#8220;When memory space is insufficient, you have to throw away what was computed and stored once and calculate it again,&#8221; Joo said, noting that &#8220;there is enormous inefficiency in the system.&#8221;<\/p>\n<p>His proposed alternative is a structure in which memory sits at the center of the system and each compute device deposits and retrieves data as needed. He described it as a shift in thinking: &#8220;If we place memory in the middle of the system and have data thrown in and taken out from there, wouldn&#8217;t far more efficient communication be possible?&#8221; He offered an analogy: rather than each household owning multiple cars, keeping one car and using shared vehicles as needed is better for resource management.<\/p>\n<p>Concretely, the picture adds a separate memory rack holding tens to hundreds of terabytes (TB) alongside GPU server racks and CPU (central processing unit) server racks. Local memory attached to existing servers stays in place while a shared pool is added on top. Racks are connected to one another via RDMA (remote direct memory access), and connections within a rack use CXL (Compute Express Link).<\/p>\n<p>The benefits were presented in experimental figures. In a multi-turn AI inference evaluation conducted with partners, the time to generate the first token averaged 80.76 seconds under the conventional recomputation method but fell to 15.45 seconds when remote CXL DRAM was used\u2014an improvement of about 5.2x. In the same experiment, loading a cache stored on a remote solid-state drive (SSD) took 33.56 seconds, giving CXL memory the speed advantage.<\/p>\n<p>Data pathAverage time to first tokenConventional recomputation80.76 secondsRemote SSD cache33.56 secondsRemote CXL DRAM15.45 seconds<\/p>\n<p>Note: Results of a multi-turn AI inference evaluation conducted by SK Hynix with partners<\/p>\n<p>Behind the problem lies a structural characteristic of conversational AI. In back-and-forth exchanges across multiple turns, the intermediate data used to process prior context\u2014the &#8220;key-value cache (KV cache)&#8221;\u2014keeps accumulating. If storage space runs short and this is discarded, the same calculations must be repeated when generating the next response. The presentation materials included an example assuming a GPT-4-class model: storing the weights requires 3.5TB, and under conditions generating 20 tokens per second, reading the weights requires 4.4TB of bandwidth per second.<\/p>\n<p>This is also why SK Hynix is focusing on Intel. &#8220;Intel has strong leadership in computing and is drawing even more attention in the agentic AI era, when CPUs are becoming more important,&#8221; Joo said, adding, &#8220;We are working to create innovation together with Intel.&#8221; The judgment is that in an environment where multiple AI agents divide roles among themselves, data exchange between agents surges and memory requirements only grow larger.<\/p>\n<p>Challenges remain before this concept can become reality. The connection specifications linking the memory pool and servers, along with the software ecosystem, must be developed together, and the latency and reliability of CXL-based shared memory must be validated at data-center scale. SK Hynix is continuing related tests with customers and partners.<\/p>\n<p>Industry observers see the announcement as noteworthy for expanding the role of memory companies. Where they previously only supplied components plugged into servers, the company now aims to intervene at the system architecture design stage and lead discussions on changing the memory hierarchy itself. Against the backdrop of AI infrastructure investment shifting from compute chips toward data movement and storage efficiency, the move is expected to serve as a gauge for the direction of demand related to memory and the CXL ecosystem. About 1,000 industry officials from South Korea and abroad attended the event, and LG Electronics also gave a presentation.<\/p>\n","protected":false},"excerpt":{"rendered":"South Korea&#8217;s SK Hynix has put front and center a so-called &#8220;memory pool&#8221; architecture in which multiple servers&hellip;\n","protected":false},"author":2,"featured_media":226291,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[20],"tags":[4560,40141,2653,106057,67430,76620,357,106059,106058,241,275],"class_list":["post-226290","post","type-post","status-publish","format-standard","has-post-thumbnail","category-sk-hynix","tag-agentic-ai","tag-cxl","tag-intel","tag-intel-ai-summit-seoul-2026","tag-joo-young-pyo","tag-kv-cache","tag-lg-electronics","tag-memory-pool","tag-rdma","tag-sk","tag-sk-hynix"],"_links":{"self":[{"href":"https:\/\/www.europesays.com\/korea\/wp-json\/wp\/v2\/posts\/226290","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.europesays.com\/korea\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.europesays.com\/korea\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/korea\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/korea\/wp-json\/wp\/v2\/comments?post=226290"}],"version-history":[{"count":0,"href":"https:\/\/www.europesays.com\/korea\/wp-json\/wp\/v2\/posts\/226290\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/korea\/wp-json\/wp\/v2\/media\/226291"}],"wp:attachment":[{"href":"https:\/\/www.europesays.com\/korea\/wp-json\/wp\/v2\/media?parent=226290"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.europesays.com\/korea\/wp-json\/wp\/v2\/categories?post=226290"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.europesays.com\/korea\/wp-json\/wp\/v2\/tags?post=226290"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}