Intel GPU D-Series Turns Crescent Island Into A 1.3 PFLOPS AI Beast
The big selling point for the D-Series is memory capacity, not speed, although it's no slouch. Intel says its GPU can accommodate models with up to 300 billion parameters on a single card, while an eight-GPU system can apparently handle models exceeding one trillion parameters. As we reported before, while the chip is technically capable of supporting up to 480GB of memory, Intel's branded PCIe card comes equipped with 160GB of LPDDR5X, while Intel's partners can develop versions with capacities of up to 480GB.
The latter configuration provides considerably more room for model weights, KV caches, and concurrent inference sessions; critically important in the era of "throw a swarm of agents at the problem." Despite being based on familiar LPDDR5X memory instead of exotic HBM, the accelerator seemingly offers a very impressive 1.5TB/s of memory bandwidth, apparently achieved using a 1280-bit memory bus and LPDDR5X running at 9.6 Gbps, making this a truly mammoth memory bus.
The D270 incorporates 32 Xe3P cores and 256 XMX matrix engines. AI-focused as it is, this silicon reportedly doesn't include any graphics functionality whatsoever. The company does emphasize broad numerical-format support, though, promising capabilities ranging from "native low-precision inference to FP64 scientific computing"; it's not clear what "low-precision inference" means, but we know from past disclosures that the part supports FP8, FP4, and microscaling formats like MXFP4.
Intel is pitching the D-Series as a relatively straightforward addition to existing data center infrastructure. The accelerator uses PCIe 5.0, operates within a 350W power envelope, and supports air-cooled deployments without requiring specialized liquid-cooling infrastructure. On the software side, Intel says the GPU is built on its common Xe software foundation and targets established AI frameworks and inference software, including PyTorch and vLLM.
This launch comes as Intel attempts to establish a more competitive position in AI acceleration following its previous Gaudi efforts that were well-regarded but not very popular. Where previous-generation AI compute hardware was often focused on model training, this newer generation often focuses specifically on serving inference, as demand for inference is reportedly exploding. So it goes that the D-Series targets inference workloads, where memory capacity, serving efficiency, and the cost of generating tokens can be as important as peak compute throughput.
Intel says customer sampling is already underway and that it is working with OEMs, ODMs, cloud providers, and software partners to bring the hardware to market. Pricing and general availability have not yet been announced; obviously, we'll need third-party benchmarks to see how Intel's specs and claims hold up against real workloads.
