Skip to content- Sources: primary, discussion
- Summary: Intel's product page lists up to 480 GB LPDDR5X per card, 32 Xe3P cores, 256 XMX matrix engines, and 350 W passive cooling on standard PCIe 5.0, with no HBM. Intel targets prefill-heavy agentic and long-context inference and names vLLM, SGLang, PyTorch, and Triton support. Performance claims are projections against RTX Pro 6000 from Intel internal analysis.
- Why it matters: It offers a high-capacity option for KV-cache-heavy serving in existing air-cooled racks, but no independent benchmarks exist yet.