Intel plans to launch its next-generation Xeon platform, codenamed Oak Stream, in 2026, which is able to embody Diamond Rapids, a CPU constructed for servers and high-performance workloads.
Diamond Rapids will use Intel’s 18A course of and Panther Cove cores, the identical structure coming to future shopper chips.
The highest mannequin contains 4 compute tiles, every with 48 efficiency cores, including as much as a complete of 192 cores per socket.
New socket kind
With assist for single-, dual-, and quad-socket configurations, which means a single rack might host as much as 768 cores. That form of density additionally brings challenges, significantly with energy. A totally configured rack might draw a whopping 2000W.
Intel says Diamond Rapids will use a brand new socket kind, LGA 9324. It additionally helps fashionable requirements like PCIe 6.0 and CXL 3, serving to the processor connect with accelerators and quick storage.
Every CPU can assist as much as 16 channels of DDR5 reminiscence and MRDIMM modules working at 12800MT/s.
There will even be a model with fewer compute tiles and a single I/O module, providing half the reminiscence bandwidth and core rely. That possibility might go well with extra power- or space-conscious deployments.
Intel can be pitching this as a CPU prepared for inference workloads. Whereas most AI processing nonetheless occurs on GPUs, Diamond Rapids goals to enhance CPU-based inference, particularly for smaller fashions.
Native assist for codecs like FP8 and TF32 ought to assist in that house.
The CPU will even assist Intel APX and an upgraded AMX engine, two instruction units meant to enhance effectivity for contemporary server duties. These upgrades trace at a future the place CPUs might tackle extra AI workloads instantly.
Diamond Rapids is predicted to launch alongside Intel’s Jaguar Shores AI accelerator, a part of its push to construct a full AI-ready server platform.
Intel hasn’t introduced pricing but, and the way it stacks up towards opponents like AMD stays to be seen.
Through expreview
