4 results about "PCI Express" patented technology

The application provides a large model training acceleration method and system based on CUDA shared memory and a PCIe expansion card, hardware configuration information is acquired to identify the PCIe expansion card and build an expansion storage space, and a unified management heterogeneous memory space is formed in combination with GPU shared memory; in the training process, according to a preset initialization strategy, model data blocks are loaded to different levels of the heterogeneous memory, data access frequency is monitored in real time, data migration instructions are dynamically generated based on multiple frequency thresholds, and intelligent scheduling of the data blocks among the levels of storage is realized; a transparent address remapping mechanism is used to enable CUDA kernel functions to seamlessly access the heterogeneous memory space; periodic data snapshots and persistent storage are simultaneously supported, and quick recovery after interruption of training is ensured; the application breaks through the GPU display memory capacity limit, improves the large model training efficiency, reduces the dependence on multi-card hardware and cost, and guarantees data security and training reliability.

A first set of parameter values are programed to a first set of sequencer registers. A second set of parameter values are programmed to a second set of sequencer registers. In response to a detecting a triggering event, a hardware sequencer performs the following operations: transfer the first set of parameter values from the first set of sequencer registers to a first set of link training registers, transfer the second set of parameter values from the second set of sequencer registers to a second set of link training registers, and initiate one end of a communication link training with a host.

A modular compute platform for artificial-intelligence inference is disclosed. Each compute module delivers exaFLOPS-class throughput within a six-hundred-watt envelope and can be installed on diverse carrier boards such as PCI Express, automotive, robotic, or rackmount systems. Modules include vertically integrated logic and memory for high-bandwidth communication and local storage of neural-network parameters. Uniform mechanical and thermal interfaces allow identical modules to serve multiple markets, enabling wide deployment of exa-scale inference capability.

A PCI Express accelerator card capable of exa-class artificial-intelligence inference within a six-hundred-watt power envelope is disclosed. The card receives twelve-volt DC power through a 12VHPWR connector and employs a multi-phase voltage-regulation network that distributes regulated power to stacked semiconductor dies on a silicon substrate. Distributed temperature sensors and predictive control firmware coordinate electrical and thermal management to maintain safe junction temperatures using standard air or liquid cooling, enabling desktop-class exa-FLOPS performance without immersion systems.