Intel Data Center Max GPU Embraces 16-pin Power Connector

Intel’s highest-performing Data Center GPU Max compute GPUs, codenamed Ponte Vecchio, will come in the form of OAM modules aimed at blade machines for supercomputing. The company is also prepping humble Data Center GPU Max add-on cards that can go into mainstream servers and workstations. Those cards will follow the latest fashion and come equipped with the 12VHPWR power connector.

Intel’s Data Center GPU Max 1100 employs a cut-down single-stack version of Intel’s Ponte Vecchio processor that features 56 Xe-HPC compute cores featuring 448 512-bit vector engines, 448 4096-bit matrix engines, 56 ray tracing cores, and 48GB of HBM2E memory. The unit consumes up to 300W of power. As it turns out, the graphics card relies on an auxiliary PCIe 5.0 power connector, according to an Intel video.

Nvidia was the first company to use the 12VHPWR power connector on a consumer graphics card, the GeForce RTX 4090. Unfortunately, for now, 12VHPWR connectors have a nasty reputation among enthusiasts. There have been 20 user reports of the 16-pin power adapter melting and three cases of the native 16-pin power connector suffering the same meltdowns. Andreas Schilling from jokingly tweeted, “Let’s hope those datacenter experts know better how to insert that 12VHPWR adapter correctly.” There has been a couple of theories from experts in the field, but nothing conclusive as we’re still waiting for the results of Nvidia’s investigation.

It is unclear why Intel used a 12VHPWR power connector for its 300W card. Still, it is possible that the company decided to adopt a PCIe Gen5 solution for its part that will be used primarily with Intel Xeon Sapphire Rapids servers designed around the new standard.

Intel’s add-in Data Center GPU Max 1100 board is designed both for mainstream servers and for workstations used by individuals who need very high AI (i.e., INT8, BF16, FP16, etc.) and HPC (i.e., FP64) performance for artificial intelligence/deep learning and various compute-intensive and simulation workloads. Using Xe-Link interconnects, the card can work as a single board, in pairs, or in 4-way configurations. Once four such accelerators are installed into a PC, they will draw 1,200W, which is by any means a lot. But with four potentially hazardous power connectors, such a system could be called unsafe.  

Currently, only Nvidia’s GeForce RTX 4090 and GeForce RTX 4080 (due to be released on November 16) use the 12VHPWR power connectors. Both AMD’s Radeon RX 7900-series and Intel’s Arc Alchemist graphics cards use standard eight-pin PCIe auxiliary power connectors.