Intel Unveils D270 Accelerator With Up to 480 GB of LPDDR5X Memory

Intel has added the D270 accelerator, with up to 480 GB of LPDDR5X memory, to its discrete data center GPU section. That is more capacity than any shipping HBM-based competitor offers: Nvidia's B200 has about 192 GB of HBM3E, and AMD's Instinct MI355X has 288 GB. The link to the product page has collected 67 points and 20 comments on Hacker News. The discussion is still small, but it keeps returning to one question: what is the bandwidth of this memory?

What is known and what is not

The only source here is a single product card in the D-series. Two facts are pulled into its headline: the device class (a GPU accelerator for AI) and the memory capacity. Everything else that usually decides whether an accelerator gets purchased is not disclosed.

• Name: Intel D270, D-series, discrete data center GPUs
• Memory: up to 480 GB of LPDDR5X
• Purpose: AI workloads, judging by the positioning
• Memory bandwidth: not specified
• TDP and cooling system: not specified
• Price and availability: not specified
• Framework and driver support: needs to be checked separately

Until Intel publishes bandwidth figures and low-precision performance numbers, it is too early to compare the D270 with competitors on speed. Only capacity can be compared.

Why LPDDR5X instead of HBM

HBM delivers terabytes per second, but it is expensive, depends on CoWoS and similar packaging processes, and remains in short supply. LPDDR5X is cheaper, produced in large volumes, and lets a vendor reach hundreds of gigabytes without complex packaging. The price is speed: with the same number of channels, bandwidth will be noticeably lower than that of an HBM stack.

For training, this trade-off is almost always a bad one. For inference, it can make sense. Token generation in large models is bound by reading the weights and the KV-cache, so both parameters matter. But with long contexts, or with several models on one device, capacity often decides whether the workload fits at all.

If the D270 continues the line Intel showed in 2025 with the Crescent Island announcement (160 GB of LPDDR5X, aimed at inference), capacity has tripled. The source does not confirm any link between the two products; this is only context.

Who could use it

The weight arithmetic is simple. In FP8, one parameter takes one byte, so 480 GB can hold a model of roughly 400-plus billion parameters, with headroom for the KV-cache and activations. At 4-bit precision, the ceiling is twice as high. On accelerators with 192 GB, such a model has to be split across at least three or four devices with tensor parallelism. A single 480 GB card simplifies the setup: fewer inter-chip links and easier deployment.

Before building the D270 into any plans, it is worth asking Intel for five things:

1. Actual memory bandwidth and FP8/FP4 performance. 2. Support for vLLM, PyTorch and popular inference servers without custom patches. 3. TDP and cooling requirements: whether an ordinary air-cooled rack will do. 4. Price per device and per token on target models, not just per card. 5. Dates for samples and volume shipments.

For now, this is a claim to the niche of cheap inference for large models with long context. Whether it holds up will depend on bandwidth and software maturity, not on the 480 GB listed on a product page.