Intel and SK Hynix outlined a collaborative approach to AI inference bottlenecks, combining CPU-led task coordination with shared memory. According to Sina Finance, Intel presented its AI technology roadmap and deployment plans at the 2026 Intel Seoul AI Summit on the 12th in Seoul, and announced the K-AI Full-Stack Alliance, which initially includes 23 South Korean companies and institutions.

Intel said CPUs can coordinate tasks among AI models and manage program execution, while GPUs handle model computations. Its Xeon 6+ server CPU, released in the second quarter, integrates up to 288 low-power cores. The in-development Crescent Island inference GPU supports multiple models; the Intel version has 160 GB of memory, while partner versions can have up to 480 GB. Intel said a single-card text-to-video demonstration had been completed, but the hardware and software were not yet ready for mass production.

SK Hynix proposed a shared memory rack that lets multiple servers access additional memory. In partner tests, generating the first token took an average of 80.76 seconds when deleted cache had to be recomputed, 33.56 seconds when cache was stored on an external SSD, and 15.45 seconds with a CXL remote DRAM solution. The latter was approximately 81% faster than recomputing the data, SK Hynix said.

The companies are also working together on memory-sharing technology and system design. Intel described applications including ultrasound image measurement without a separate GPU; the company said inference throughput was 4.2 times that of the previous solution.