Intel D270 GPU leaks: 480GB memory, 1.5TB/s bandwidth, built for local LLM inference
karminski3 · x · 2026-10-15
Intel's new Crescent Island D270 GPU, shown at the OCP summit, targets high-density local LLM inference:
- Native OCP MXFP4/MXFP8 support (no NVFP4); 1.3 PFLOPS FP8, 2.6 PFLOPS FP4.
- Retail cards cap at 160GB (480GB is ODM-only), both with LPDDR5x at 1.5TB/s over a 1280-bit bus.
- Theoretical: 111 tps decode on a 4-bit 27B model, 187 tps on DeepSeek-style sparse 4-bit; 48K tps prefill for 27B.
- Weakness: no NVLink-class interconnect, PCIe 5.0 only (64GB/s) — fine for inference and 70B LoRA fine-tuning, not for large-scale training.
- vLLM, SGLang, and PyTorch already supported; rumored $20K pricing.
More from Infra
- Google expands Android Earthquake Alerts from 98 to nearly 200 countries after 1B+ alerts — sundarpichai · 2026-10-16
- Microsoft's Magnetar zkVM proves programs 10-39x faster using only standard hash functions — Microsoft Research · 2026-10-16
- Moving 1PB back on-premises costs a fortune as AI training drives cloud repatriation — DavidLinthicum · 2026-10-15
- FP8 + DSpark stack delivers ~2x speedup for Nemotron 3.5 Super VL on H100s — vllm_project · 2026-10-15
- Nemotron 3.5 Super VL hits ~630 tokens/s at 100k-token prompts, fastest 20+ score model measured — ArtificialAnlys · 2026-10-15
- vLLM Ships Day-0 Support for Nemotron 3.5 Super VL with NVFP4 on a Single GB200 — vllm_project · 2026-10-15