Intel says customer sampling is underway for its D270 data-center GPU, an inference accelerator announced at the Open Compute Project Global Summit. Intel's release lists 160 GB of LPDDR5X on its branded PCIe card and says partner designs can be configured with up to 480 GB; the D270 product page lists the 480 GB maximum per card.
A generic PCIe graphics card, not Intel's D270. Photo: Drgulcu / Wikimedia Commons, CC BY-SA 3.0.
That distinction matters for long-context inference. Model weights share accelerator memory with a cache that stores context for generating the next token; the cache grows as more text and concurrent requests remain active. FP8 is an 8-bit number format, so a simple one-byte-per-weight estimate puts 300 billion weights at about 300 GB before that cache and other runtime needs. Intel says the D270 can support models up to 300 billion parameters on one card, but it does not identify the memory configuration or numerical precision behind that ceiling. The estimate is arithmetic, not a D270 test.
Intel lists 1.5 terabytes per second of memory bandwidth and 1.3 petaflops of FP8 compute within a 350-watt, air-cooled design. Its product page specifies PCIe 5.0. Those are company specifications; Intel has not published an independent benchmark in the announcement, and the release gives no price or on-sale date. Customer sampling is underway, so the product is not yet presented as broadly available.
The 480 GB option could keep larger models or more context in memory, but capacity alone does not establish token speed, concurrent throughput or total deployment cost. Buyers will need configuration-specific results that state precision, model, context length and workload before comparing it with other inference systems.
Sources
- Intel's OCP 2026 announcement, including the 160 GB branded-card footnote and sampling status.
- Intel GPU D-series / D270 product page.
- Image source and license record.