Intel GPU D270 Targets AI Inference and HPC with 160 GB Memory at 350W

Intel GPU D270 Targets Financial Modeling and AI Inference at 350W

Intel has detailed the GPU D270 accelerator, a 350W PCIe 5.0 card built on the Xe 3P architecture with 160 GB of memory. Formerly known as Crescent Island, the chip combines low precision AI math with full rate FP64 compute for scientific modeling and financial workloads. Intel is demonstrating pre production hardware at SC26 in Chicago, though retail pricing and release timing remain unannounced.

A total of 32 Xe cores and 256 XMX engines sit inside the air cooled PCIe form factor. Each core executes 64 FP64 fused multiply add operations per cycle alongside FP4 and FP8 math for inference. The standard Intel card carries 160 GB of LPDDR5X memory at 1.5 TB/s bandwidth, while partner boards scale up to 480 GB. Keeping large datasets on the GPU avoids slow transfers across the host bus. Intel also shared pre production numbers comparing the 350W D270 against the 600W NVIDIA RTX PRO 6000 Blackwell Server Edition. The D270 won on options pricing despite drawing much less power. Throughput jumped 5.9x in European Monte Carlo, 5.6x in Black Scholes, 4.6x in Binomial, and 1.92x in American Monte Carlo workloads.

Running AI models inside scientific simulations creates nasty hardware bottlenecks. Intel splits that work across specialized chips instead of forcing everything onto 1 processor. Under a live SC26 setup, the D270 runs the compute heavy prefill stage for large simulation context windows. SambaNova SN50 Reconfigurable Dataflow Units handle the memory bound decode stage, while Intel Xeon 6 and Xeon 6+ CPUs coordinate the workflow. On standalone AI tests against the NVIDIA RTX PRO 6000 Blackwell card, the D270 cut latency by up to 29% on the Drex 2.0 decision model. It also hit 1.12x higher throughput running Qwen 3 30B A3B.

Hardware means nothing if developers have to rewrite their entire codebase. Intel backs the card with Intel MPI, Fortran, C++, standard AI frameworks, and open SYCL tools so existing software runs from day 0. Viridien proved this by porting a 3D seismic imaging app from HIP and CUDA over to SYCL for the SC26 demo. The team finished the port in under 1 month without changing host code. Almost 90% of the compute kernels ran well without Intel specific tuning, and complex kernels reached up to 70% of theoretical peak speed right away.

About the author

mgtid
mgtid
Owner of Technetbook | 10+ Years of Expertise in Technology | Seasoned Writer, Designer, and Programmer | Specialist in In-Depth Tech Reviews and Industry Insights | Passionate about Driving Innovation and Educating the Tech Community Technetbook

Join the conversation

Newsletter Subscription