Scalable Link Interface

from Wikipedia, the free encyclopedia

Scalable Link Interface , SLI for short, is a multi-GPU technology from Nvidia that enables two or more graphics chips to be interconnected to increase rendering performance (SLI Frame Rendering) or to use up to four screens (SLI Multi View).

Theoretically, depending on the operating mode, the rendering performance corresponds on average to the computing power of the slowest graphics processor times the number of graphics processors used. With two structurally identical graphics cards, the computing power is theoretically doubled, in practice, however, performance increases of around 30 to 90% are possible - in some applications, however, the computing power is lower than with a single graphics card.

history

The term SLI as Scan Line Interleave first appeared in 1998 at the former graphics chip manufacturer 3dfx , which enabled several of its graphics chips ( Voodoo Graphics , Voodoo 2 and VSA-100 ) to be coupled . Nvidia SLI was only introduced in 2004. Except for the abbreviation and the idea of ​​distributing the processing load to two or more graphics chips under SLI Frame Rendering, the two methods have nothing in common, although the term SLI continued to be used after Nvidia took over 3dfx.

Before Nvidia SLI published, tried the graphics chip maker ATI already with Alternate Frame Rendering . Today's improved multi-GPU technology from ATI is called " Crossfire " and differs in part from Nvidia's SLI technology. In 2003, the manufacturer XGI also presented two graphics cards with multi-GPU technology.

With the introduction of the Turing CUDA core generation in semiconductors and graphics cards (RTX models), Nvidia finally swapped SLI technology for the newer and much more powerful NVLink technology. However, the corresponding bus signals are usually only found in the top models, as it typically makes more sense to fully exhaust the technical possibilities on a board before attempting a system with multiple graphics semiconductors. This successor has currently only been implemented in the mass market for systems with exactly two cards that are connected with a bridge. Only high-performance workstations, such as models from the DGX series or computing nodes with Power + main processors, allow e.g. B. with the SXM / SXM2 module standard (instead of PCI-E) for the graphics semiconductor a significantly greater NVLink connectivity.

requirements

Requirements for using SLI are, on the one hand, a motherboard with an SLI-compatible chipset and, on the other hand, several SLI-compatible graphics cards, be it as onboard graphics or via additional PCI-E slots. So far, NVIDIA has largely limited the use of SLI to its own chipsets, which also have to bear the abbreviation SLI. Only the Nehalem X58 chipset announced by Intel can be activated for SLI by the manufacturer against payment of a license fee to NVIDIA. There are no technical reasons for this limitation - it is a pure marketing decision on the part of NVIDIA. There are modified drivers that enable SLI to be used on other chipsets as well. These are mostly out of date and there are currently no efforts to activate SLI-compatible drivers of the latest generation, as new drivers appear too regularly and the effort is not worth it.

Initially, a special SLI bridge (in German "bridge") was necessary to connect the cards, so that only graphics cards with these connections could be connected to one another. In the meantime, NVIDIA has also made SLI possible for cards without this connection, the necessary data exchange between the cards is handled via PCI Express. Since this bus is limited in bandwidth, this type of SLI is only useful with relatively slow graphics cards; faster cards should still be better connected via the SLI bridge.

SLI only works with identical graphics chips. Sometimes it even has to be exactly the same graphics card with an identical BIOS version. With newer drivers , however, the SLI function becomes increasingly flexible. It should also be noted that up to Nvidia driver version 180.xx, only one monitor output worked when SLI frame rendering was activated . Several monitors can now be used with active SLI frame rendering .

Problem of micro-stuttering

Both Crossfire and SLI mainly use alternate frame rendering . At the moment, however, the phenomenon often occurs that every second frame is calculated much longer than the previous one. This leads to an uneven game flow, which is particularly annoying at relatively low frame rates and is often referred to as "micro stuttering". Due to this fact, cards in SLI mode (including dual GPU cards) require significantly higher average frame rates than a single card in the games in question, in order to still guarantee a relatively smooth display for the eye.

The problem of micro-stuttering has been alleviated with the latest drivers. The frame, which is calculated very quickly, is slightly delayed when it is output to the screen by means of an intermediate buffer. The frame rates become more even again and the maximum latencies between two consecutive frames are reduced, as the waiting time for the following, slowly calculated frame is reduced. With this workaround, there are significantly fewer micro-stutters and the games become more playable, even if the average FPS number drops slightly. The delay time used, however, artificially increases the latency time from the initiation of the image calculation to its display for the faster calculated image.

Operating modes

SLI knows the following operating modes:

  • SLI frame rendering:
  • SLI Multi View:
    • Multi View for controlling 4 and more monitors. (Graphics cards work independently of one another in this mode.)
  • Nvidia Surround (from GeForce GTX 260 )
    • Surround technology. When using two or three GPUs in SLI mode, more than two monitors (3 monitors) can be controlled. In contrast to the multi-view approach, there is no need to do without SLI frame rendering.
    • 3D vision surround. When using two to three GPUs in SLI mode (also with DUAL GPU cards such as GTX-590), three monitors can be used to output stereoscopic image content.

The following table provides an overview of the maximum theoretical bandwidth for data transfer depending on the type of bridge, as can be found on the free market:

Clock rate of the bridge 400 MHz 650 MHz
Simple bridge 1 GB / s 1.625 GB / s
Double bridge 2 GB / s 3.25 GB / s

SLI graphics chips

The following graphics chips support SLI:

You still need graphics cards with the same graphics chip, but no longer from the same manufacturer (exception: Quadro-FX series).

Dual GPU graphics cards

In addition, there are graphics cards that are already equipped with two graphics chips and therefore already run in SLI mode.

A multi-GPU system can also be implemented with just one slot for graphics cards on the motherboard, with two slots then Quad-SLI is possible.

On June 6, 2006, NVIDIA introduced the first multi-GPU card "7950 GX2" for the mass market. The “2” at the end of the name indicates the two 7900GT GPUs; However, these run well below the reference clock of a single 7900GT card, as the two existing heat sinks would be too small for the otherwise generated waste heat.

In March 2008 nVidia again presented a dual GPU graphics card under the name Geforce 9800 GX2 . The name was not continued in the 200 series, but the dual GPU card of the series is the Geforce GTX 295 . nVidia relied on the same design principle and installed the GPUs on two interconnected boards. Instead of two coolers, this time a common cooler was installed, which is located between the two circuit boards. Corresponding recesses in the boards ensure the air supply to the fan.

Thanks to its two GPUs, the dual GPU card achieves a higher performance than previous single cards, as it has SLI practically from the factory. However, the power consumption is correspondingly high and the efficiency is significantly lower than with a single card.

Quad SLI

If you connect two Nvidia dual GPU graphics cards (exceptions: Asus EN7800GT DUAL, Gigabyte Dual 6600GT), you get a quadruple SLI system (Quad SLI). However, the increase in performance remains rather disappointing, especially when you consider the high costs for such a configuration. At CeBIT 2006 the first performance-enhancing, stable quad SLI systems were on display.

3-way SLI

3-way SLI from GeForce 8800 Ultra graphics cards

3-way SLI (Triple-SLI) enables the connection of three Nvidia GeForce graphics cards of the same model with 3-way SLI support via a special SLI bridge in one system. The first chipsets to support 3-way SLI are the nVIDIA nForce 680i SLI, nForce 780i SLI, nForce 780a SLI, the nForce 790i SLI and the nForce 790i Ultra SLI.

The following graphics cards support 3-way SLI:

4-way SLI

In addition to Quad-SLI (consisting of two dual GPU graphics cards), there is also 4-way SLI, but this should not be confused with Quad-SLI. 4-way SLI is implemented by four Nvidia GeForce graphics cards of the same model with 4-way SLI support, a 4-way SLI-supporting motherboard and a special SLI bridge. Because four PCI Express x16 slots are required on the motherboard, but not enough PCI Express lanes are available for the chipset , so-called "bridge chips" are often used with 4-way SLI. These bridge chips (e.g. Nvidia NF200) help ensure that enough lanes are available and that no performance should be lost.

The following graphics cards support 4-way SLI:

Hybrid SLI

With hybrid SLI, it is possible to switch on the GPU in the chipset with a low-cost graphics card and then operate it in SLI mode. This feature is called "GeForce Boost" by nVidia. In addition, the chipset GPU can be switched off completely in 3D mode (this is necessary with stronger graphics cards, otherwise the chipset GPU would slow down the external graphics card) and only let the external graphics card render. In Windows operation, the chipset GPU is then switched back to and the external graphics card is deactivated. NVIDIA calls this process “HybridPower”. The first chipsets to support hybrid SLI are the 730a, 750a SLI and 780a SLI for the AMD platform. However, the graphics card must have an SMBus in order to be able to be switched off by the HybridPower control . The first graphics cards to support HybridPower are the GeForce 9800 GX2 and Geforce 9800 GTX . GeForce Boost is supported by the low-cost graphics cards of the GeForce 8 series , Geforce 9 series and Nvidia Geforce 9M series . Hybrid-Power was only developed for Windows Vista up to the beta phase, but was then based on the reason that the effort for testing the compatibility of the respective circuit boards with the graphics cards was too high and therefore the costs were not inconsistent with the rather restrained customer interest calculate, set. Since Nvidia is withdrawing further and further from the chipset business and the two large x86 CPU manufacturers are integrating the GPU into the CPU, the "GeForce Boost" technology is no longer being developed any further.

A modification of Nvidia's “HybridPower” technology and Intel's “Switchable Graphics” is available for mobile graphics chips, the so-called “Optimus” technology . It switches back and forth between the dedicated GPU from Nvidia and Intel's integrated graphics chip, depending on the load. The commands to switch are given by Nvidia's graphics driver, which Nvidia provides with a list of 3D applications, based on which the driver switches to the dedicated Nvidia GPU when such an application is started. This function is only available under Windows 7 and newer Windows operating systems.

Individual evidence

  1. ↑ NVIDIA's SLI performance: double the power for youngsters at tomshardware.de
  2. ↑ SLI for PCs without Nvidia chipset at Heise.de
  3. ↑ The multi-GPU problem (micro stuttering) Computerbase, January 19, 2008, accessed February 16, 2011.
  4. ↑ ComputerBase : The problem "micro stuttering" Computerbase, January 24, 2011, accessed on February 16, 2011.
  5. ↑ Computerbase: Frame Pacing AMD vs. Nvidia , August 1, 2013, accessed January 5, 2014
  6. ↑ https://www.pcgameshardware.de/Pascal-Codename-265448/News/HB-SLI-Bruecke-Untnahme-1204297/
  7. ↑ GeForce GTX 465, GTX 470, GTX 480 4-way SLI support from 197.75 Nvidia.de
  8. ↑ GeForce GTX 285 and GTX 295 without hybrid SLI Nvidia provides information about the deleted hybrid power feature

Web links