Intel Xeon 6716P-B socket: BGA 4368

What the Xeon 6716P-B is

The Intel Xeon 6716P-B is the 40-core flagship of the Xeon 6 Granite Rapids-D soldered edge stack on BGA 4368. It runs 40 Performance-cores with Hyper-Threading for 80 threads at 2.5 GHz base with boost to 3.5 GHz, backed by 160 MB of cache and a 235 W TDP. Intel lists the part at a 3584 USD recommended price, built on Intel 3 for single-socket edge servers, with listings spanning the first two quarters of 2025 across Intel pages.

Forty cores suit dense virtualization hosts, storage clusters, data firewalls and inference nodes that consolidate many services onto one soldered board. The 160 MB cache pool is the largest in the BGA 4368 range, which helps large working sets stay near the cores. At 235 W the part needs high-airflow edge chassis with vendor-qualified heatsinks.

Against the 40-core Intel Xeon 6706P-B beside it, the 6716P-B shares clocks and cache as a companion bin for different board qualifications. Against the 36-core Intel Xeon 6553P-B below it, the 6716P-B adds 4 cores for denser consolidation. Zero UPI links confirm single-socket-only topologies. See Socket vs chipset: what decides whether a CPU works for how socket, platform and firmware fit together on server boards.

Socket and motherboard compatibility

The 6716P-B fits BGA 4368, a 4,368-ball soldered package measuring 77.5 × 50 mm. The chip is soldered once at the factory and stays with the board for life, so there is no retention mechanism and no aftermarket mounting choice. Cooling is part of the board design for the 235 W class with an 85 degree case limit.

Board choice follows the server vendor list for Granite Rapids-D edge systems. Memory is 4-channel DDR5 ECC fixed per board, and expansion rides on 32 Gen 5 plus 16 Gen 4 PCIe lanes. A board revision needs firmware that knows the Granite Rapids-D stepping. Moving from any other socket is a full platform change with board, memory and heatsinks replaced together.

Memory, PCIe and cooling

Memory on BGA 4368 is DDR5 ECC in 4 channels, up to 6400 MT/s on qualified boards. Four channels keep 80 threads supplied in virtualization, storage and inference mixes. Some boards qualify at 5600 MT/s instead, so the vendor memory list governs exact speeds and population rules.

PCI Express connectivity is Gen 5 with x32 lanes plus Gen 4 with x16 lanes from the CPU. That lane budget feeds NVMe arrays, 100-gigabit networking and accelerators on a dense single-socket node. AVX-512 with two FMA ports and DL Boost accelerate inference where the workload uses them.

Cooling follows the board vendor list for the 235 W class with vendor-qualified heatsinks. No cooler ships with the tray server part. Firmware power limits can cap draw below the default where a cabinet demands it, at some cost in sustained clocks.

Upgrade paths

On the same package, there is no larger BGA 4368 part above the 40-core pair in this batch, so growth on BGA 4368 means adding nodes rather than a bigger chip. The 36-core Intel Xeon 6553P-B below it keeps high clocks at the same 235 W class.

The step down to the 36-core Intel Xeon 6556P-B trims cores for the 215 W class. A move to BGA 5026 is a different soldered platform with 64 cores and 8 memory channels for denser nodes. That change replaces board, memory and cooling from scratch.