← sources

Intel confirms 1.5 TB/s and 1.3 PFLOPS FP8 for its 480 GB LPDDR5X air-cooled inference card

source: reddit (reddit:stocks)published 2026-10-13 08:18ingested 2026-10-14 02:40

retained text

At the Open Compute Project Global Summit on 12 October 2026 Intel released technical details for its data-centre GPU D-series. The card reaches up to 480 GB of LPDDR5X memory, 1.5 TB/s of memory bandwidth and 1.3 PFLOPS of FP8 compute inside a 350 W air-cooled standard PCIe card. Customer sampling is underway. Intel's own branded card carries 160 GB while partner ODM cards reach 480 GB. Intel says it supports models up to 300 billion parameters on a single card and more than 1 trillion parameters across an eight-GPU system.



Here's what I think this implies.



Intel designed the D-series and is the only vendor sampling it, so whether capacity-bound inference buyers take the card accrues directly to its data-centre accelerator line.



NVIDIA is exposed the other way. Intel benchmarks the card against the RTX PRO 6000, the incumbent air-cooled PCIe inference card with a fraction of the memory, so NVIDIA loses capacity-bound serving share if a 350 W card holding a 300-billion-parameter model proves adequate.



The earliest test is Intel's third-quarter results on 29 October 2026. The question is whether Intel attaches D-series customer commitments, volume-production or general-availability timing to the sampling it has now confirmed. Separately, the first ODM announcement of a 480 GB D-series card would confirm the high-capacity configuration is actually being built.



If published or independently measured serving throughput on long-context, high-concurrency workloads lands far below HBM-based alternatives at the same model size, the capacity advantage will not convert into deployments. Intel has released peak FP8 compute and bandwidth but no throughput figures. Equally, if the 480 GB configuration never ships as an ODM part and the line stays at Intel's 160 GB card, the single-card 300-billion-parameter claim never reaches customers.



Source:  
[https://www.intel.com/content/www/us/en/products/details/discrete-gpus/data-center-gpu/d-series.html](https://www.intel.com/content/www/us/en/products/details/discrete-gpus/data-center-gpu/d-series.html)



Happy to be wrong on the NVIDIA side.

document fetched 2026-10-14 00:33 · canonical https://reddit.com/r/stocks/comments/1x4r0ij/intel_confirms_15_tbs_and_13_pflops_fp8_for_its