Intel disclosed new GPU D-Series details on October 12, 2026, at OCP 2026. Formerly called Crescent Island, this family targets AI inference in data centers, emphasizing memory capacity and air-cooled deployment.
The announced specifications
Intel lists Xe3P architecture, a 350W power envelope, 1.5TB/s memory bandwidth and 1.3 petaflops of FP8 compute. These are manufacturer specifications, not BYT01 measurements or an independently established performance ranking.
Memory capacity and documentation limits
Intel’s current product page and announcement state up to 480GB of LPDDR5X. Neither current page corroborates the previously attributed 160GB footnote, so that distinction has been removed from our summary. Confirm the partner configuration and exact card before assuming an installed capacity. Intel says customer sampling is underway.
Decision and evidence limits
BYT01 conclusion: This platform concerns data-center inference; start with the model, software and service requirements.
| Item | Evidence boundary |
|---|---|
| Article type | Manufacturer news and specifications; not a gaming-card recommendation. |
| Material limitation | Intel advertises up to 480GB. The two current official pages do not corroborate the previously attributed 160GB footnote. |
| Measurement requirements | Compare model precision, context, concurrency, framework, latency and measured power; peak FP8 alone is insufficient. |
Cost comparisons require matched configurations, workloads and latency targets, including hardware, energy, software and service costs. An announcement is not a supply quote or a demonstrated business case.
This editorial revision used AI assistance and has no recorded independent human expert sign-off.
Test protocols and evidence · Commercial disclosure · Corrections policy · Related decision guide
Update — 23 October 2026: Revised the decision summary, evidence limits, comparison conditions, disclosure and links. This is an editorial revision date, not a new hardware measurement or a recheck of every retailer offer.
BYT01’s perspective
Evaluate an inference accelerator against the actual model, context length and required software. A peak FP8 figure does not establish application speed or deployment cost. BYT01 has not verified local pricing or availability, and this announcement is not a gaming-PC purchase recommendation.
For the CPU side of this workflow, read our coverage of AMD Helion’s agentic AI rack design.
Sources and verification
BYT01 checked these sources on October 14, 2026. This is original news coverage, with manufacturer figures and plans attributed to their publishers. BYT01 has not tested these products.