🇰🇷South Korea·AI News·12 Nov 2026·via 서울경제

    Intel to Sell AI Inference GPU as Standalone Chip, Tripling DRAM Capacity

    Intel will offer its next-generation AI inference GPU, Crescent Island, as a standalone chip, allowing server manufacturers to configure graphics cards with up to 480GB of DRAM. This capacity is three times Intel’s own reference design and aims to address varying memory requirements for different AI inference workloads, as outlined at the Intel AI Summit Seoul 2026.

    Intel's decision to supply its Crescent Island GPU as a standalone component provides server makers with flexibility in designing their AI systems. This approach allows customers to tailor the amount of DRAM, specifically LPDDR5X low-power memory, to match the demands of particular AI services, contrasting with a fixed configuration for greater customization.

    The need for flexible DRAM capacity stems from how AI inference systems handle data. When an AI analyzes long documents or maintains conversations, it stores previously processed information in a key-value (KV) cache within the DRAM. Insufficient DRAM capacity forces the system to recompute discarded KV cache data, which can slow response times. By enabling up to 480GB of DRAM, Intel aims to help optimize performance and response for diverse AI applications.

    This strategy could also affect the semiconductor supply chain. By adopting LPDDR5X DRAM for Crescent Island, Intel is expected to broaden supply opportunities for major memory manufacturers like Samsung Electronics and SK hynix. Both companies are already mass-producing low-power DRAM suitable for AI servers, positioning them to benefit from Intel's entry into the inference GPU market.

    Share this article

    Go deeper
    Original reporting by 서울경제We don't republish, read the full story →

    Related reading

    6 stories