Intel details D270 inference GPU with up to 480 GB LPDDR5X and 350W air cooling
- Intel's D270 is an inference-optimized data center GPU with up to 480 GB of LPDDR5X memory per card, a 350W air-cooled design on standard PCIe 5.0.
- The card has 32 Xe3P cores and 256 XMX matrix engines, and Intel says die area was moved from graphics and ray-tracing pipelines to inference compute.
- Intel pitches the D270 as an alternative to HBM for long-context agentic workloads, keeping more of the model and context resident so fewer GPUs are needed.
- Intel claims zero-day support for leading open models on vLLM and the open PyTorch ecosystem, with no proprietary networking required.
- The page gives no price or availability date, and Hacker News commenters pointed out that its FAQ still contains placeholder text.
Hacker News 의견들
Can I actually buy this? Not now, and maybe not ever. The board shape says it's meant for a server chassis behind 80mm blower fans, but you'll probably find one used on eBay in 7 years for a home lab.
Seems like very little compute for that much memory. 1.5 TB/s and 1.3 PFLOPS with no fast multi-GPU link means it's basically a consumer card dressed up as a server card.
That's the point, I think. Fast local 2 to 4 user service on very large MoE models is the niche, and banks with in-house servers might buy these for experiments.
They originally said 120GB per card, and partners could ship bigger capacities. So much for that. 20 channels of LPDDR5 is still one of the few ways to get that much memory for cheap, though.
Fun to see the LLMism on the product page: "A, not B" phrasing everywhere.
Is there any indicator of what this will cost per unit? The FAQ on that page is clearly placeholder text, so I don't trust it on specs either.
Given the rumored numbers hold, this could be a real contender against the M5 Ultra. Low-batch single-user serving is mostly bottlenecked by memory bandwidth times memory size per tier.
Honestly these AI hardware announcements read like Maybach commercials to me. Not targeting me at all.