Xcena Demonstrates CXL-based AI Inference Memory Solution at OCP Summit
Published2026.10.16 09:28
KV 캐시는 AI가 이전에 처리한 정보를 저장해 답변 생성 시 재사용하는 데이터로, 서버마다 개별 저장하는 대신 CXL 메모리 풀에 집중 저장함으로써 중복 저장과 서버 간 데이터 이동을 줄일 수 있다고 설명했다.
또한 MX1의 데이터 인접 프로세싱(NDP) 기능을 활용해 검색증강생성(RAG) 과정 중 벡터 검색을 가속하는 데모도 함께 선보였다.
데이터를 호스트 프로세서로 반복 전송하는 대신 데이터가 저장된 메모리에서 직접 연산을 수행하는 방식이다.
양사는 MX1의 SSD 메모리 확장 기능 ‘인피니트메모리(InfiniteMemory®)’를 활용해 기존 CMM-D 대비 메모리 용량 확장성을 높이면서도 SSD 대비 나은 성능을 구현할 수 있음을 확인했으며, 실제 AI 서빙 환경에서의 적용 가능성을 검증 중이다.
SK하이닉스는 행사 현장에서 MX1 칩을 활용한 ‘CXL 하이브리드 메모리’와 RDMA 기반 아키텍처를 공개했다.
15일에는 삼성전자와 에이전틱 AI용 분리형 CXL 메모리 아키텍처를 공동 발표했다.
그랜트 맥키(Grant Mackey) 수석 엔지니어는 15일 마이크로소프트·씨게이트·메타 연사들과 OCP 컴포저블 메모리 시스템(CMS) 기술 업데이트 세션에 참여해 메모리 용량·대역폭 대응 현황을 공유했다.
Comments
Related news
Xcena and Portrai Join Forces to Build CXL-Based Bio Data Infrastructure
Fabless semiconductor company Xcena and techbio company Portrai are embarking on joint development of large-scale biodata processing infrastructure based on CXL. They are building a solution that processes rapidly increasing analysis data driven by advances in spatial omics technology, transcending existing DRAM limitations through CXL memory expansion. The two companies have completed pre-validation of data at the 35 million cell scale and plan to expand to a maximum of 100 million cells while applying Xcena's next-generation solution 'MX1'.
2026.10.06by 배종인 기자
Xcena Recruits Steve Fingerhut, Former Qualcomm and Intel Executive, as Chief Business Officer
Semiconductor fabless company Xcena has recruited veteran Steve Fingerhut from Qualcomm and Intel as Chief Business Officer. With over 25 years of experience in data center and memory business, he will lead the commercialization of CXL-based compute memory solution 'MX1' and expansion into global markets. Xcena is focused on applying MX1 to AI infrastructure of hyperscalers and cloud service companies.
2026.10.14by 배종인 기자
IBS Implements Semiconductor, Insulator, and Electrode in One Step through Atomic Substitution…Develops Next-Generation 2D Semiconductor Process
A research team at the Institute for Basic Science (IBS) has successfully developed a technology to selectively replace atoms in 2D semiconductors and achieved fabrication of 'modular transistors' that simultaneously implement semiconductors, insulators, and semimetals on a single thin film. The researchers substituted selenium atoms in WSe₂ with oxygen, sulfur, and tellurium to form an insulator (WOx), a semiconductor (WS₂), and a semimetal (WTe₂), respectively, while also implementing atomic-level precision control and selective processing. The fabricated devices demonstrated low energy barriers and excellent current control performance, and implementation of CMOS inverters confirmed feasibility for integrated circuit applications. This achievement is expected to contribute to the development of ultra-low-power, ultra-fine semiconductors and three-dimensional stacked semiconductor technology.
2026.10.16by 배종인 기자
ASML records record-breaking Q3 results... "ASML production capacity determines the ceiling for advanced semiconductor supply"
ASML's Q3 2026 results demonstrate that the semiconductor industry is entering a new phase. The company achieved record-breaking results with revenue of 11.662 billion euros and net profit of 4.96 billion euros, but what captured market attention was the warning that "it will be difficult to meet customer demand even with production capacity planned for 2027-2028." As AI proliferation has caused advanced semiconductor demand to surge, the industry's bottleneck is shifting from insufficient demand to inadequate EUV equipment supply. With expanded memory investment and accelerated High-NA EUV adoption underway, future semiconductor market competitiveness is expected to be determined by securing production capacity rather than price.
2026.10.16by 배종인 기자
Arm Announces CPU Benchmarking Methodology for Agentic AI
Arm has announced a CPU performance evaluation methodology specialized for agentic AI workloads. Developed with support from Fujitsu, Meta, and OpenAI, this methodology focuses on performance measurement based on actual agent task completion results. According to independent verification by Signal65, the Arm AGI CPU demonstrated an average performance advantage of 1.99x per rack compared to AMD EPYC Turin, with performance differences of 1.7x to 2.2x particularly in the tooling, orchestration, and inference domains.
2026.10.15by 배종인 기자
