Intel D270 GPU targets NVIDIA with 160GB memory and FP64 compute

Intel D270 is aiming NVIDIA with some bold speed claims. The Crescent Island GPU targets AI and heavy science tasks, with Intel claiming gains of up to 5.9 times in select tests against the RTX PRO 6000 Blackwell Server Edition. But those results come from Intel itself, not outside labs.

The hardware has a fair bit going on. Its Xe 3P design packs 32 Xe cores and 256 XMX engines. Math support spans FP4 through FP64, with each core able to do 64 FP64 fused multiply-add operations. That last bit matters for science, finance, and engineering work where high precision is a must.

RAM is another big part of the pitch. Intel fits 160GB of LPDDR5X onto its base card, with 1.5 TB/s of bandwidth. Partners can build versions with as much as 480GB. That gives teams room to run large data sets and AI models without as much need to move data off the GPU.

Power use looks quite tame for this class of gear. The D270 has a 350 W cap, uses PCIe 5.0, and relies on air cooling. That could make it easier to fit into server racks that lack liquid cooling.

Intel also claims up to 29% less lag in the Drex 2.0 test. For Qwen 3 30B-A3B, it cites a 1.12 times throughput edge at 350 W. Those are neat numbers, but the test setups vary, and Intel used chips that were not yet final.

Meanwhile, the NVIDIA card in the Qwen test had 96GB of GDDR7 and a 600 W power cap. Intel's card had far more RAM and a lower power limit. That makes the result worth a look, though it does not prove the same edge will hold in other apps or AI tasks.

On the software front, the D270 works with SYCL, oneAPI, and common HPC tools. Intel also has a plan for split AI inference, where its GPU handles the prompt stage, SambaNova gear runs the decode work, and Xeon CPUs manage the flow. That setup is optional, not a rule for use.

For labs and data centers, the FP64 math and large RAM pool may be the main draw. Yet Intel still has to prove that its tools work well across real tasks and that it can ship enough cards. There is no firm sale date or public price for the D270, and no sign of a home PC version.
 

Attachments

  • Intel D270 GPU targets NVIDIA with 160GB memory and FP64 compute.webp
    Intel D270 GPU targets NVIDIA with 160GB memory and FP64 compute.webp
    16.9 KB · Views: 1

Sponsored

Top