Intel’s D270 puts memory at the centre of its AI inference pitch

Intel’s D270 is a data-centre accelerator designed to run AI models, with memory capacity aimed at long inputs and many simultaneous users. Customer sampling is under way, but Intel has not announced general availability or pricing.

Intel AI Watch analysis

What happened

The D270 is part of Intel’s GPU D-series, previously known as Crescent Island. The Gadgeteer reports that Intel shared further details at the Open Compute Project Global Summit on 12 October. Intel’s listed uses include enterprise assistants, document processing, coding assistants and image-generation models.

Intel specifies 160GB of memory on its own PCIe card; manufacturing partners could build cards with up to 480GB, according to a footnote in the company’s announcement. Intel says a single card can support models with up to 300 billion parameters, or more than a trillion across eight GPUs. Those are capacity claims, not a measure of how quickly a model will answer.

The card uses PCIe 5.0, has a 350-watt power envelope and is passively air-cooled, with airflow supplied by the server. Intel names PyTorch and vLLM among the supported software. The Gadgeteer’s account says customer sampling is under way; it gives no general-availability date or price.

Why it matters

Long documents, extended conversations and busy services all need memory to hold model data and the information being processed. Intel is pitching the D270’s capacity and conventional server cooling as a way to accommodate that workload without requiring specialised liquid cooling. The useful question for operators is not simply how large a model fits, but how many requests the complete system can serve at an acceptable speed.

Our read

There is a clear, practical design pitch here: more memory, ordinary server airflow and familiar AI software. But the largest numbers do not tell operators what they most need to know about performance. Before anyone gets out the procurement paperwork, the tests that matter are workload-specific throughput, latency, power and software compatibility.

What to watch

• When Intel plans to make the D270 generally available, and at what price.
• Independent performance and power results on real inference workloads.
• Which memory configurations, models and software stacks customers actually sample.

Discussion spark: For inference hardware, should buyers prioritise the amount of model memory a card can hold, or wait for independent evidence of speed and cost per request?

Independent WittyWires tracker for public updates about Intel AI. Not affiliated with or endorsed by Intel AI; this is not an official account.