Intel GPU D270

The Intel GPU D270 is a dedicated data center accelerator engineered specifically for large-scale AI inference. Representing the debut of Intel's GPU D-Series (previously code-named Crescent Island), the D270 departs from conventional GPU conventions by stripping out legacy 3D graphics hardware, rasterization units, and ray-tracing pipelines. Every square millimeter of its silicon is committed to compute and high-density memory handling, establishing the card as an enterprise accelerator built to tackle the memory-capacity bottlenecks that stall large language model deployment.

Available in a standard dual-slot PCIe 5.0 x16 form factor with a 350W thermal design power (TDP), the D270 runs on passive air cooling. This design choice bypasses the expensive liquid-cooling loops and specialized infrastructure modifications required by many contemporary accelerators. The card ships in two distinct hardware profiles: an Intel-branded reference configuration carrying 160GB of memory, and partner board designs scaling up to a massive 480GB of LPDDR5X memory.

By prioritizing dense LPDDR5X capacity over raw High Bandwidth Memory (HBM) pin-speed, Intel targets real-world production environments where model footprints and massive key-value (KV) caches dictate hardware requirements. For engineers evaluating the best hardware for running AI models locally or across on-prem cluster nodes, the D270 shifts the economics of running 70B to 300B parameter models on minimal PCIe slots.

Evaluating the Intel GPU D270 for AI workloads requires looking at how its architecture balances memory capacity against compute density. Built on the Xe3P architecture, the D270 features 32 Xe3P execution cores and 256 Intel Xe Matrix Extensions (XMX) engines backed by 32MB of unified L2 cache.

• Architecture: Intel Xe3P (Crescent Island)
• Compute Cores: 32 Xe3P cores
• Matrix Acceleration: 256 XMX engines
• Dense Compute: 1.3 PFLOPS FP8
• Memory Capacity: 160GB LPDDR5X (Intel reference) / up to 480GB LPDDR5X (partner designs)
• Memory Bandwidth: 1.5 TB/s (1500 GB/s)
• Cache: 32MB unified L2
• Interconnect: PCIe 5.0 x16
• TDP: 350W
• Cooling: Passive, air-cooled server chassis integration

In LLM serving, token generation is memory-bandwidth bound during the decode phase. At 1.5 TB/s, the D270 matches or exceeds typical GDDR6-based cards, though it trails the 3.0 TB/s to 5.3 TB/s seen on flagship HBM3e accelerators. However, the use of LPDDR5X allows a single card to hold up to 480GB of memory. This solves the primary failure mode of local LLM serving: running out of VRAM when allocating memory for expanded context windows and multi-user concurrency.

The card delivers 1.3 PFLOPS of dense FP8 throughput. The 256 XMX engines handle matrix multiplication directly in 8-bit precision, enabling high-throughput prompt evaluation (prefill phase) where operations are compute-bound rather than memory-bound. At a 350W envelope, the D270 achieves strong compute density per watt, allowing data center operators to populate standard 4U and 8U air-cooled chassis without exceeding standard rack power limits.

The practical value of a 480GB GPU for AI is the elimination of multi-node tensor parallelism for large model tiers. A single D270 partner card can fit models up to 300 billion parameters that traditionally required four to eight discrete accelerators.

• Llama 3.1 8B & Mistral 7B: These models require less than 18GB in native 16-bit precision. On the D270, smaller models can be served with hundreds of thousands of concurrent tokens in the KV cache, or deployed in multi-instance configurations serving dozens of isolated agent streams on a single card.
• Llama 3.1 70B & Qwen 2.5 72B: At 16-bit weights, a 70B model requires roughly 140GB of VRAM. It fits entirely onto the 160GB Intel-branded card with overhead for context. On the 480GB partner variant, a 70B model can run unquantized with full 128k context lengths across massive batch sizes without offloading or KV cache paging to host RAM.
• DeepSeek-V2.5 & Dense 300B Parameter Models: At FP8 precision, a 300B model occupies roughly 300GB to 320GB of memory. The 480GB partner card holds the entire model in memory on a single PCIe slot, leaving more than 100GB of spare capacity for KV cache allocations.
• Llama 3.1 405B: The flagship open-weights model fits on a single 480GB card when quantized to INT4 or FP4 (requiring approximately 230GB to 250GB for weights). For FP8 execution, two D270 cards provide up to 960GB of combined memory pool over PCIe 5.0, accommodating the full-precision 405B model alongside long context windows.
• DeepSeek-R1 (671B MoE): With active parameters hovering around 37B per token, an eight-card cluster of D270 accelerators provides up to 3.84TB of unified memory, easily fitting trillion-parameter mixture-of-experts architectures without exotic inter-chassis networking.

The architecture is tuned for FP8 execution. While INT4 provides higher absolute generation speed, FP8 maintains near-lossless accuracy on modern architectures like Qwen 2.5 and Llama 3.1 while fully engaging the 1.3 PFLOPS XMX matrix engines.

In single-stream generation, Intel GPU D270 tokens per second correlate with memory bandwidth:

• 70B (FP8): Generates approximately 18 to 22 tokens per second for single queries.
• 70B (INT4): Generates approximately 32 to 38 tokens per second.
• Batched Throughput: In high-concurrency environments, where multiple requests are processed in parallel, the compute pipelines saturate, driving aggregate throughput up into hundreds of tokens per second across active streams.

The Intel GPU D270 addresses specific enterprise and production requirements rather than home desktop setups.

Agentic architectures generate extended prompt histories, multi-turn tool outputs, and autonomous reasoning loops that blow past normal VRAM budgets. The D270 is among the best hardware for local AI agents, offering the capacity needed to keep multi-agent contexts hot in memory without context truncation or continuous re-summarization.

Organizations running private AI inference behind corporate firewalls can deploy the D270 directly into existing standard 19-inch server racks. Because the card uses standard passive heat sinks and a PCIe 5.0 form factor, IT departments do not need to replace power distribution units (PDUs) or plumb liquid cooling lines.

• Local Desktop Hobbyists: As an air-cooled enterprise PCIe card requiring forced chassis airflow, the D270 is not designed for consumer desktop cases or low-noise environments.
• Frontier Pre-Training: While the D270 can perform lightweight fine-tuning (LoRA, QLoRA) on large models, it is not an alternative to NVIDIA H100/B200 clusters for pre-training massive foundation models, as it lacks high-speed proprietary inter-node fabrics like NVLink.

To position the D270 within the current hardware landscape, it helps to compare it against alternative PCIe enterprise options.

The NVIDIA L40S is a popular PCIe-based enterprise inference card based on the Ada Lovelace architecture.

• Memory Capacity: The D270 outclasses the L40S significantly, offering up to 480GB compared to the fixed 48GB GDDR6 on the L40S.
• Bandwidth: The D270 delivers 1.5 TB/s, whereas the L40S provides 864 GB/s.
• Model Support: Running a 70B model on the L40S requires quantization down to 4-bit or using a multi-GPU array. The D270 runs 70B models at FP16 on a single card with memory to spare, and accommodates 300B models on partner boards.

The H100 NVL is NVIDIA's dedicated dual-GPU PCIe inference card.

• Compute and Bandwidth: The H100 NVL provides higher raw bandwidth (nearly 3.9 TB/s per card using HBM3) and higher raw compute, giving it an advantage in latency-critical, single-stream generation speeds.
• Memory Density: The H100 NVL tops out at 94GB of memory per GPU (188GB total). The 480GB D270 partner design provides more than double the memory capacity of a full H100 NVL dual-card setup, making the D270 the more cost-effective option for running large models that are bottlenecked by capacity rather than raw compute cycles.

The Intel GPU D270 delivers an effective, high-density solution for teams looking for an alternative to fragmented multi-GPU setups, offering the capacity to serve 300B parameter models within a manageable 350W air-cooled envelope.