For GPU and CPU processors in accelerated systems, NVLink is a high-speed link that drives data and computations to useful outcomes. Once limited to high-performance computers in government research facilities, accelerated computing has become widely available.
AI supercomputers are being used by banks, automakers, factories, hospitals, merchants, and others to handle the increasing amounts of data that they must process and comprehend.
These strong, effective systems are computing superhighways. On their lightning-fast trip to actionable answers, they transport calculations and data via parallel routes.
The resources along the route are CPU and GPU processors, and their onramps are quick connections. NVLink is the industry standard for accelerated computing interconnects.
What is NVLink?
A reliable software protocol creates the high-speed link between GPUs and CPUs known as NVLink, which usually runs on many wire pairs printed on a computer board. It enables lightning-fast data transmission and reception across processors from shared memory pools.
At speeds of up to 900 gigabytes per second (GB/s), NVLink, which is now in its fourth iteration, links host and accelerated processors.
The bandwidth of PCIe Gen 5, the link found in traditional x86 servers, is more than seven times that amount. Additionally, because NVLink uses just 1.3 picojoules per bit for data transfers, it has five times the energy efficiency of PCIe Gen 5.
History of NVLink
With each successive NVIDIA GPU architecture, NVLink has improved in tandem since its initial release as a GPU connection with the NVIDIA P100 GPU.
When NVLink first connected the GPUs and CPUs in two of the most potent supercomputers in the world, Summit and Sierra, in 2018, it gained significant attention in the field of high performance computing.
Installed in Oak Ridge and Lawrence Livermore National Laboratories, the systems are advancing science in areas like drug development and catastrophe forecasting, among others.
Bandwidth Doubles, Then Grows Again
The third-generation NVLink, which came with a dozen interconnects in each NVIDIA A100 Tensor Core GPU, increased the maximum capacity per GPU to 600GB/s in 2020.
AI supercomputers in cloud computing services, business data centers, and HPC laboratories worldwide are powered by the A100.
One NVIDIA H100 Tensor Core GPU now contains eighteen fourth-generation NVLink interconnects. Additionally, the technology has assumed a new, strategic function that will allow for the world’s most sophisticated CPUs and accelerators.
A Chip-to-Chip Link
A board-level connection called NVIDIA NVLink-C2C is used to combine two processors onto a single device, forming a superchip. For instance, the NVIDIA Grace CPU Superchip, a processor designed to provide energy-efficient performance for cloud, corporate, and HPC customers, combines two CPU chips to produce 144 Arm Neoverse V2 cores.
A Grace CPU and a Hopper GPU are additionally joined via NVIDIA NVLink-C2C to form the Grace Hopper Superchip. It combines accelerated computation for the most demanding AI and HPC tasks on a single chip.
One of the first to employ Grace Hopper will be Alps, an AI supercomputer slated for the Swiss National Computing Center. The high-performance supercomputer will tackle huge research challenges in domains ranging from quantum chemistry to astrophysics when it goes online later this year.
Additionally, Grace and Grace Hopper are excellent in reducing energy consumption in demanding cloud computing tasks.
The Grace Hopper processor, for instance, is perfect for recommender systems. In order to provide billions of users with trillions of results every day, these internet economic engines require quick and effective access to large amounts of data.
A potent system-on-chip for automakers that incorporates NVIDIA Hopper, Grace, and Ada Lovelace processors and uses NVLink. NVIDIA DRIVE Thor is a vehicle computer that integrates cognitive features including entertainment, automatic driving, parking, digital instrument cluster, and more into a single architecture.
LEGO Links of Computing
NVLink functions similarly to the socket that is imprinted onto a LEGO component. It serves as the foundation for constructing supersystems to handle the most challenging AI and HPC tasks.
For instance, an NVIDIA DGX system’s eight GPUs’ NVLinks exchange quick, direct connections using NVSwitch chips. When combined, they allow for an NVLink network in which each server’s GPU functions as a single system.
DGX workstations itself may be stacked into modular units of 32 servers to provide even higher performance, forming a strong, effective computing cluster.
Using an NVLink network within the DGX and an NVIDIA Quantum-2 switched InfiniBand fabric between them, users may integrate a modular block of 32 DGX devices into a single AI supercomputer. An NVIDIA DGX H100 SuperPOD, for instance, has 256 H100 GPUs to provide up to an exaflop of the best AI performance.
Users may access cloud-based AI supercomputers, like the one Microsoft Azure is constructing with tens of thousands of A100 and H100 GPUs, to achieve even higher performance. Some of the biggest generative AI models in the world are trained using this service by organizations like OpenAI.
Additionally, it serves as another illustration of the potential of accelerated computing.
