Source
How XPUs Meet a World-Class AI Factory
NVIDIA introduced NVLink Fusion to connect custom accelerators, known as XPUs, with its established AI infrastructure. This allows companies to build AI factories by combining custom hardware with proven networking and rack-scale architecture.
Why it matters
This allows companies to bring custom AI chips to market faster and with less financial risk. They can use standardized power and cooling instead of designing these complex systems from scratch.
The details
- Sixth-generation NVLink has 3x lower latency than off-the-shelf Ethernet for XPU transfers.
- NVLink-C2C is 6x more energy efficient than PCIe when connecting XPUs to CPUs.
- The system allows mixing GPUs and custom XPUs using MGX rack-scale architecture.
Show entities and relationshipsHide entities and relationships
In this article
Products
Topics
Key connections
NVIDIA owns NVIDIA GB300 NVL72
NVIDIA designs and builds the GB300 NVL72 rack-scale system.
NVIDIA owns NVIDIA B300
NVIDIA manufactures the B300 GPU accelerator.
NVIDIA owns NVIDIA NVLink-C2C
NVIDIA provides NVLink-C2C for chip-to-chip interconnect connectivity.
NVIDIA owns NVIDIA MGX
NVIDIA created the MGX modular rack-scale architecture.
NVIDIA owns NVIDIA Omniverse DSX AI Factory Blueprint
NVIDIA developed the Omniverse DSX AI Factory Blueprint for digital twin modeling.
NVIDIA owns NVIDIA NCCL
NVIDIA develops the NCCL collective communications library.
Show 34 more connectionsShow fewer connections
NVIDIA owns NVIDIA NIXL
NVIDIA develops NIXL software for AI factory infrastructure.
Annapurna Labs was acquired by Amazon
Annapurna Labs is an Amazon subsidiary company.
QCT is a member of Quanta Computer
QCT is a data center solutions division and subsidiary of Quanta Computer.
Jack Luoh serves as head of product and solution at QCT.
Jack Luoh works at Quanta Computer
Jack Luoh serves as head of product and solution at Quanta Computer.
Tim Wilson works at Intel
Tim Wilson is VP and general manager of data center silicon engineering at Intel.
Vince Hu is corporate senior VP and general manager at MediaTek.
Lie-Szu Juang leads GUC
Lie-Szu Juang serves as chair and chief strategy officer at GUC.
CC Lee works at Annapurna Labs
CC Lee is senior hardware development manager at Annapurna Labs.
Intel collaborates with NVIDIA on CPU integration for the NVLink Fusion ecosystem.
MediaTek is a partner of NVIDIA
MediaTek partners with NVIDIA on the NVLink Fusion program for data center computing.
GUC partners with NVIDIA to integrate custom ASICs and XPUs via NVLink Fusion.
QCT collaborates with NVIDIA on system manufacturing and rack integration leveraging NVLink Fusion.
Quanta Computer is a partner of NVIDIA
Quanta Computer manufactures hardware leveraging NVLink Fusion and MGX architectures.
Annapurna Labs is a partner of NVIDIA
Annapurna Labs uses the NVL72 rack design and supplier ecosystem enabled by NVLink Fusion.
NVLink Fusion uses NVIDIA NVLink
NVLink Fusion brings custom XPUs into the NVIDIA NVLink scale-up domain.
NVLink Fusion uses NVIDIA NVLink-C2C
NVLink Fusion includes NVLink-C2C for connecting XPUs to host and ecosystem CPUs.
NVLink Fusion is related to NVIDIA MGX
NVLink Fusion adopters can utilize the NVIDIA MGX rack-scale architecture and supply chain.
NVIDIA GB300 NVL72 uses NVIDIA NVLink
NVIDIA GB300 NVL72 uses a 72-GPU NVLink scale-up domain for high throughput.
NVIDIA GB300 NVL72 competes with NVIDIA B300
GB300 NVL72 delivers higher per-GPU throughput and interactivity compared with NVIDIA B300.
NVIDIA Vera Rubin NVL72 uses NVIDIA Vera CPU
NVIDIA Vera Rubin NVL72 integrates Vera CPUs into its rack-scale architecture.
NVIDIA Vera Rubin NVL72 is built with NVIDIA MGX
NVIDIA Vera Rubin NVL72 is built using the NVIDIA MGX rack-scale architecture.
NVIDIA Omniverse DSX AI Factory Blueprint is related to NVLink Fusion
The blueprint provides digital twin modeling for gigawatt-scale AI factories using NVLink Fusion.
NVIDIA NCCL is related to NVLink Fusion
NCCL provides collective communications for distributed workloads across NVLink Fusion factories.
NVIDIA Dynamo is related to NVLink Fusion
Dynamo enables infrastructure disaggregation in NVLink Fusion AI factories.
NVIDIA NIXL is related to NVLink Fusion
NIXL provides software support for disaggregated AI factory infrastructure.
NVIDIA Mission Control is related to NVLink Fusion
Mission Control handles cluster management, telemetry, and debugging for AI factories.
NVIDIA NVLink-C2C is related to PCIe
NVLink-C2C delivers up to 6x the energy efficiency of a PCIe interface.
NVIDIA NVLink competes with Ethernet
NVLink provides 3x lower latency and 10x higher packet rates compared to off-the-shelf Ethernet.
NVIDIA NVLink uses Co-Packaged Optics
Future NVLink roadmap configurations incorporate co-packaged optics across large accelerator domains.
NVIDIA Vera Rubin NVL72 uses Liquid Cooling
Reference trays and NVLink switch trays feature 100% liquid cooling for continuous operation.
NVLink Fusion is related to Mixture of Experts
NVLink Fusion accelerates scale-up networking for mixture-of-experts model architectures.
NVLink Fusion is related to Agentic AI
NVLink Fusion removes control-to-compute barriers for agentic AI workloads.
NVIDIA Omniverse DSX AI Factory Blueprint uses Digital Twin
Omniverse DSX provides a digital twin open reference design for gigawatt-scale AI factories.
Related events
NVIDIA Introduces NVLink Fusion for Custom XPUs in AI Factories
Get the weekly recap
The stories like this one, picked and explained — once a week, straight to your inbox.