One of the most important developments in recent years has been that of external intercommunication interfaces for communicating several processors with each other. The most important being AMD’s Infinity Fabric and NVIDIA NVLink. The differences between the two? The NVLink cannot be found on the PC and it works differently. Read on to find out what the differences are.
How does NVIDIA NVLink work?
NVLink is different from PCI Express although it is used for the same. While PCI Express is the classic cross interconnect, in the case of NVLink we are dealing with a network interface and so it is the implementation of a NoC, which relies on the transmission of packets of the same. way that in a network.
Each transmitted packet is made up of what NVIDIA calls Flits, one Flit is 128 bits of data, and there are 18 Flits per data packet. The first Flits being responsible for configuring where each packet goes and how the different elements communicate with the NVLink interface.
Whichever version we are talking about, the NVLink is based on 8 full duplex lines per interface, which means that each interconnect cable is used for both sending and receiving data. Each NVLink external interface has what NVIDIA calls links, which is 4 in version 1.0 and 6 in version 2.0 of the interface.
Unlike what happens with PCI Express where each interface communicates in one direction in both directions, in the case of NVLink the interconnection takes place at what are called links, allowing up to 4 elements to interconnect with each other in the case of version 1.0 of the standard and 6 in version 2.0.
The origins of NVLink are in SLI
Originally invented by 3Dfx as Scan Line Interleave for its 3Dfx, the original SLI was based on rendering a scene by having each scan line in the image rendered by one of the two 3D maps. This evolved years later to ScanLine Interconnect to connect two GPUs via alternate image rendering, in which each GPU renders an alternate image to the other.
But SLI came with huge limitations, one of them was the inconsistency between the two GPUs as each had its own separate VRAM from the other, which not only required VRAM to be duplicated, but also only two completely equal cards could be joined, forcing symmetrical graphics card configurations to communicate with each other.
Because applications are generally intended to communicate with a single GPU control processor, the SLI works in such a way that the first GPU dominates the second, so that while the subordinate GPU generates the final frame, the final frame must be copied via a DMA mechanism to the VRAM memory of the first GPU, which is the one with the video output.
NVIDIA NVLink as a replacement for PCIe?
In the PC, we use the PCI Express port to communicate the CPU with the GPU, NVIDIA instead uses the NVLink to communicate the CPU and GPU in some systems instead of PCI Express as is the case of its Drive PX based on its NVIDIA Tegra where the automotive SoCs are directly associated with a dedicated NVIDIA GPU.
Another example is IBM POWER9 CPU supercomputers, which have an NVLink interface implemented to communicate directly with NVIDIA Volta GPUs. So he didn’t use PCI Express interfaces either.
We also know that the combination of the future NVIDIA Grace supercomputer processor with NVIDIA GPUs will be carried out using version 4.0 of the standard, so that thanks to this NVIDIA can have an exclusive ecosystem thanks to its proprietary interconnection.
But for NVLink to be standardized to replace PCI Express, PC processors would have to have adopted this interface, which has not happened and most of its advantages over PCI Express are already provided by both Compute Express Link. and by AMD Infinity Fabric.
If NVIDIA had its own x86 processor, we would probably have seen NVLink implemented on the PC and NVIDIA would have released pairs of its hypothetical x86 processors and GPUs.
Why is it better than PCI Express?
The first generation NVLink has a transfer rate of 20 Gb / s per address and link, allowing it to achieve 160 Gb / s of aggregate bandwidth. Version 2.0 has 25 Gb / s bandwidth per address and link, but its 6 units per interface allow it to achieve 300 Gb / s of aggregate bandwidth.
Even NVLink version 1.0 is higher in bandwidth than PCI Express, which in theory would mean higher power consumption, but the NVLink interface, unlike PCI Express, was not created to power cards. graphics and its power consumption is very high. Express only 5.5W per link.
Therefore, in environments where every watt of consumption is important such as in the world of supercomputers where a few years ago the ratio of all system energy consumption to total computing power was counted as a measure of performance, l The use of a more efficient interface not only in communication but also in energy is much better.
Your future goes through optical interfaces
NVIDIA has already announced its interest in creating an optical version of NVLink. Taking advantage of one of the main advantages of this type of interconnection is the fact that the signal does not degrade, allowing GPUs to intercommunicate with each other at a distance of up to 100 meters. The other advantage? The consumption per transmitted bit goes from 8 Pj / bit to 4 Pj / bit allowing configurations up to 600 Gb / s under the same consumption and allowing an increase in the number of links per NVLink interface.