Nvidia’s DGX B300 is designed for LLM inference and training and was first announced back at GTC 2025, will contain a brand new purpose built Intel Xeon 6 CPU inside.  The 64-core Xeon 6776P will be paired with the 16 Blackwell Ultra GPUs found in each DGX B300, and will incorporate two new features, priority core turbo (PCT) and speed select technology turbo frequency (SST-TF).  These new features will limit most of the CPU cores to their base frequency, allowing the non-limited cores to maintain their upper peak frequency more consistently.  This should both lower energy consumption and improve performance, as LLM models don’t tend to need every single core running at their absolute fastest.

Intel tried this in Alder Lake, but that predated their E-core and P-core design, so this new Granite Rapids family should be more effective than their predecessors.  The relationship between Intel and NVIDIA has extended through several generations of DGX systems, it’s been about five years since an AMD chip was used.  This is good news for Intel, who are not doing well against AMD in the server market.

Intel has something to worry about though, NVIDIA is looking add ARM support to NVLink Fusion so they can take advantage of the 1.8 TBps NVLink interconnect in systems with Nvidia graphics.  Nvidia are also adding support for third-party AI accelerators to work with their Grace CPUs, giving them yet another way to corner the AI market.