Intel GPU D-Series: inference card with up to 480GB LPDDR5X now sampling

What happened and what to do

The article reports that Intel's D-Series GPU, formerly Crescent Island, offers up to 480GB of LPDDR5X at 1.5TB/s and 1.3 PFLOPS in FP8, on a 350W PCIe card now sampling. Intel's own card has 160GB, while partners reach 480GB. According to the text, the large memory on a 350W card targets LLM inference capacity and cost per token in serving.

For a company, the practical point is memory capacity: large models and long contexts are often limited by memory before compute. Before any purchase decision, it is worth measuring the real workload, with the business's own models, context sizes and concurrency, and comparing cost per token and latency against current hardware. It is also possible to build a benchmark and per-request cost monitoring pipeline, with private inference on the client's infrastructure, so the decision rests on data once the platform is commercially available.

How the consultancy can help

Inference capacity and cost diagnosis: we map your models, context lengths and request volume, and compare cost per token and latency across configurations. Then we implement the benchmark pipeline and private inference in your environment, with continuous monitoring.

Next step

If your company is evaluating hosting language models on its own infrastructure, send a short description of your case, with models, volume and constraints. We will reply with a scoped proposal.