Accelerated computing solutions that were once limited to high-performance computers in government research labs are now mainstream. Banks, car manufacturers, factories, hospitals, retailers and companies from many industries are using supercomputers to cope with their increasing computing power requirements.
These systems, which are quite powerful and work efficiently, use various hardware, including GPU and CPU. So much so that advanced technologies are required for high efficiency and data transfer. Here is a solution that comes into play at this point. The gold standard in interconnects for accelerated computing: NVLink.
This technology was not actually used only in the high-performance computing class. The idea of interconnecting multiple video cards has existed since the late 1990s. While the idea sounded good, “SLI technology” never lived up to expectations. But not backing down, NVIDIA created NVLink as the successor to SLI. Even if not used by end consumers in games, these connection techniques give life to many supercomputers.
We will talk about the possibilities that NVLink offers shortly. To summarize, this high-bandwidth connection technology is now only used in business systems.
What is NVLink?
Used in accelerated systems, NVLink is a high-speed connection technology built by an advanced software protocol for GPUs and CPUs, typically running on multiple pairs of wires printed on a computer card. With this technology, processors can send and receive data very quickly from shared memory pools.
The fourth generation of NVLink is currently in use. With all the improvements made, the mainframe and accelerated processors can transfer data at speeds up to 900 gigabytes per second (GB/s). If we make a comparison, speeds of more than 7 times the bandwidth reached by PCIe Gen 5, the interconnect used in traditional x86 servers, can be achieved. NVLink increases the energy efficiency of PCIe Gen 5 by 5 times, thanks to data transfers that consume only 1.3 picojoules per bit.
SLI is Out of Time
We all know the SLI technology, which allows us to use two identical NVIDIA graphics cards in one system. The green team supported NVLink on the RTX 2070 SUPER, 2080 SUPER, and RTX 2080 Ti models. However, when we consider the cost and performance gains and take an overview of the table, we always say that SLI technology is illogical when it comes to gaming. NVIDIA only added NVLink support to the GeForce RTX 3090 in the RTX 3000 series based on the Ampere architecture.
Later, RTX 4000 series graphics cards with Ada Lovelace architecture were released. As you can imagine, the NVLink connector on the cards is completely obsolete. This step means that multi-GPU support and SLI (via a separate physical interface) are completely gone.
NVIDIA CEO Jensen Huang says they’re making room for a different feature in the I/O connection interface. So they didn’t want to waste resources and time using the NVLink interface. NVIDIA engineers also made the most of their silicon space. In this context, it is said that each span of space is used for “more powerful artificial intelligence processing”.
Difference Between NVLink and SLI
Unlike SLI, NVLink uses mesh networking, a local mesh topology where infrastructure nodes are directly connected in a non-hierarchical manner. This allows each node to transmit information instead of routing everything through a specific node. NVLink does not need a master-slave relationship as in SLI; one card in the installation works as the “master”.
SLI used Alternate Frame Rendering to perform frame rendering, meaning that each connected card renders different frames. In a dual GPU setup, one card was processing even frames while the other was processing odd frames. While this was a sensible solution, it didn’t work well in practice (mainly due to hardware limitations) and caused frustrating hangs.
For faster image processing, NVLink Bridge (NVLink Bridge) is a better solution. SLI bridges had a bandwidth of 2 GB/s at best, but the NVLink Bridge reached speeds of up to 200 GB/s in extreme cases. With the improvements for computing, the bandwidth increased up to 900 GB/s as we said.
The 160 and 200 GB/s NVLink bridges were only used on NVIDIA’s professional grade Quadro GPUs. Bridges for consumer GPUs are slower but still way better than SLI. It was possible to connect two Titan RTX or two RTX 2080 Ti graphics cards with 100GB/s bandwidth.
Chip-to-Chip Connection
NVIDIA NVLink-C2C is a version that provides board-level interconnection to create a superchip by combining two processors in a single package. For example, the NVIDIA Grace CPU Superchip, designed to deliver energy-efficient performance for cloud, enterprise and HPC users, connects two chips together. In this way, Grace can contain 144 Arm Neoverse V2 cores.
NVIDIA NVLink-C2C also combines a Grace CPU and a Hopper GPU to create the Grace Hopper Superchip. Thus, processors designed for high-end HPC and AI workloads come together in a single solution.
Alps, an artificial intelligence supercomputer planned for the Swiss National Computing Center, will be one of the first systems to use Grace Hopper. Once the supercomputer kicks in, it will work on big science problems in fields ranging from astrophysics to quantum chemistry.
Additionally, NVLink is used in a powerful system-on-a-chip for automakers that includes NVIDIA Hopper, Grace and Ada Lovelace processors. NVIDIA DRIVE Thor is an on-board computer that combines intelligent functions such as digital instrument cluster, infotainment, autonomous driving, parking and more in a single architecture.
