What does a Video Card do, what affects the performance?

We have to realize that the data as soon as it leaves the CPU has to go through 4 steps until it finally reaches the monitor:

As you can see, except the step from the RAM DAC to the monitor, each step is some kind of a bottleneck and crucial for the overall performance of the graphical subsystem. The slowest step is the one which determines the overall speed. Lets now discuss, what these single steps mean and what actually happens:

The transfer of data between CPU and the Video Chipset

This bottleneck is mainly depending on the bus type and speed, the mainboard and its chipset. The fastest bus system at present is the PCI bus, so you will have slower performance with VL bus, ISA, EISA and NuBus (only for Macs). The PCI bus however doesn't always run at highspeed of 33 MHz, so with a Pentium 75, P90, P120, P150 you'll have a PCI bus speed of only 25 (P75) or 30 Mhz, which already here decreases the performance of the graphical subsystem. Later chipsets also offer faster PCI performance, so the Intel 430HX chipset offers a faster PCI performance than the Intel Triton 430FX chipset. As the name already says, it's not a bus, it's a port. This means you can only run one device on it, the graphic device. It can runs at 66 MHz and can tranfer data at the rising and falling edge of a clock circle (x2 mode). This makes it at least double as fast as PCI, but this does not necessarily result in double performance of AGP graphic cards, because the data transfer bandwidth is not the limiting factor of current graphic cards.

The transfer of data between Video chipset to Video RAM and from Video RAM to the RAM DAC

I have been taking these two steps together because here lies the key for the performance of a video card as long as you exclude special chipset features. The big problem of a video card is that the poor video memory lies in between two very busy devices and has to serve both of them all the time. Each time the screen has to change the chipsets has to alter the video memory (and it changes continuously, e.g. mouse pointer, cursor blinking, etc.). Also the RAM DAC has to read out the video memory continuously, to maintain the screen. You can see, the video memory is caught in between them and here all these smart ideas like using VRAM, WRAM, MDRAM, SGRAM, EDO RAM, or increasing the video bus size like 32 bit, 64 bit and now 128 bit come in.

The higher the screen resolution and the higher the colour resolution, the more data has to be transferred from the video chipset to the video memory and the faster the data has to be read by the RAM DAC to be sent to the monitor. You can see that the video memory has to be accessed all the time by the chipset and the RAM DAC. Normal dynamic RAM can only be accessed at a max. frequency, so after the video chipset finished accessing (r/w) the video memory, the RAM DAC has to wait until it's allowed to read and vice versa.

The Video Card Manufacturers found 3 different ways to fight that problem:

1. Here comes the idea in, to make the video RAM dual ported. This means, that the video chipset reads or writes from/to the video memory via one port, but the RAM DAC reads out the video memory through an independent second port. The video chipset doesn't have to wait for the RAM DAC anymore and the RAM DAC doesn't have to wait for the video chipset anymore. This kind of video memory is called VRAM. It's obviously more complicated by having double the ports and therefore more expensive to produce. That's the simple reason why VRAM cards are more expensive and also faster. The WRAM used by Matrox and a few other cards is also dual ported, but organised somehow smarter so that it's faster than VRAM but also 20% cheaper to produce. If you should wonder why typically cards which offer a high refresh rate and high colour depth have these two kinds of memory, you should consider the following. A higher refresh rate means that the RAM DAC feeds the monitor with a complete screen picture more often than at a lower refresh rate. Therefore the RAM DAC has to read out the video memory more often. This only can be achieved with either VRAM/WRAM, by accessing the video memory via the second port, or by a considerable decrease of video performance of DRAM/EDO cards. If you don't believe it, just run your favourite video benchmark at a low and then at a high refresh rate - you'll see a considerable difference if you've got a DRAM/EDO card. The same is valid for a higher colour depth. At a 8 bit colour resolution (=256 colours) a 1024x768 screen needs 786,432 bytes to be read by the RAM DAC to send a complete screen picture to the monitor. At 24 bit colour resolution (16,777,216 colours) the same screen needs 2,359,296 bytes to be read by the RAM DAC - and this takes more time. This btw is also the reason why you often can't have the same high refresh rate at true colour as you had at low colour in cheaper cards.

2.  The other way to fight this problem is to increase the video memory bus size. Years back everybody was amazed by the new 32 bit video cards. These cards had a 32 bit data path between video chipset, video memory and RAM DAC. With 32 bit data path you can transfer 4 bytes in one go. Later there came the 64 bit video cards = 8 byte in one go, which are the standard at present and only recently some new chipsets were born, to have a 128 bit data path = 16 byte in one go. It's easy to see, that video cards with both (VRAM/WRAM & wide data path) will be the best performers, but with a really wide data path you could get around VRAM/WRAM. Now by getting completely excited about these wide data paths we shouldn't forget one very important thing: a normal 8x1Mbit memory chip, as used on most video cards has a data bus of 32 bits !!! Therefore even a 128 bit chipset can access this memory chip only 32 bit wide !!! This is the reason why all 64 bit video cards are a lot slower if only fitted with 1 MB of video RAM ! Don't get a 64 bit video card with less than 2 MB !!!! Chipsets with 128 bit data path usually need at least 4MB local memory, otherwise their performance is cut in half. The NVidia Riva chipset e.g. is able to talk to only 2 MB as well, via a 64 bit data path. Riva cards with only 2 MB are therefore castrated. However, due to the architecture of the card you won't use 128 bit data path even if you upgrade to 4 MB, because the data path just stays the same. This is probably the case in many video cards, so be careful not getting a 1 MB 64 bit card or a 2 MB 128 bit card!

3. The third and to us maybe most obvious way to get the video RAM accessed faster is to simply increase the clock speed of the video chipset/video RAM/RAM DAC. Years back the video chipsets ran at clock speeds high above the mainboard memory bus speeds already. SGRAM is nowadys running at 100 MHz clock and some graphci chip manufacturers are already talking of 125 or even 133 MHz video RAM clock using 7 ns SGRAM. SGRAM is nothing but a special graphics version of SDRAM (synchronous DRAM), so we know this is able to run at clock speeds up to 133 MHz.

Summarizing all these performance aspects, we learn that for optimal performance we should have an AGP or at least PCI system with the latest chipset and 33 MHz PCI bus speed, a video card with a high performance chip and either SGRAM or WRAM, a wide data path or a high clock frequency of the video chipset or best all these three things together!

Now which parts of this video card/monitor-combo plays which role?

The video card, which is inevitable in each computer system, is responsible to process the special video data received from the CPU into a format that a monitor can understand to make a rastered picture on the screen out of it. More or less the monitor screen is still the output device of a computer system, it's the most important port through which we humans get data transferred from the computer. Therefore the Video Card/Monitor-Combo is one of the most important parts of our computer and we should take very good care of it.

Now what do we have to ask of this Video Card/Monitor Combo ?

Picture Quality is very important, so that it's not gonna be a pain for our eyes to sit and look at the screen. Here are the factors for picture quality:

Speed of the video device. Due to the work we do with the computer, the picture changes eversooften and these changes should take place as fast as possible. It depends on what in particular we do on the computer, so there are also some catagories:

Considerations for Gamers

In case gaming is most that you do on your system and you couldn't care less for Windows NT, true color and OpenGL, you want to go for a pure 3D gaming card or get an add on card.

Considerations for Professionals

The introduction of real 3D-accelerators started with the 3Dfx Voodoo chip, and that was the time when the first hard core 3D gamers started to evolve, producing high expectations as well as crazy hypes and some pathological fanaticism too. Voodoo (1) and 3Dfx remained at the top of the 3D graphics scene for more than a year, completely underestimated by the big players in the graphics scene at this time, S3 and Matrox. 3Dfx won the second round also, the Voodoo2 was a worthy successor of Voodoo (1). The idea of Voodoo2 was the same as for Voodoo (1), offer an add-on 3D-card to the normal 2D card, use a pass-through cable and offer a lot more power than Voodoo1. Voodoo2 was using two parallel texture units, a lot more graphics memory and the user had the chance to run two of those cards in parallel, called 'SLI'-mode, offering at that time mind blowing 3D performance. This was certainly no cheap solution, but the hard core gamers jumped on this bandwagon right away, making Voodoo2 another huge success. It took six more months until the first decent 2D/3D-cards became available, NVIDIA shipped RIVA TNT and 3Dfx tried their slightly castrated version called 'Voodoo Banshee'. People who believed in the marketing hype, saying that those two chips were a really new generation of 3D chips, were disappointed, since Voodoo2 SLI was still offering the best performance, Banshee and TNT had only the chance to come close in some areas.

Voodoo3, TNT2, Permedia3 and what not are the first 3D-chips that show a real further development of the technologies implemented into Voodoo2. What took no less than six chips in the days of Voodoo2 SLI is now condensed into a TNT2-chip, beefed up with a faster RAMDAC, good AGP-support, support of larger texture-sizes and several other goodies, running at a much higher clock speed than Voodoo2 as well. Really new concepts are still not to be found though and can also not be expected before the second half of 1999. We are still working with the second generation of 3D-chips, but even this generation has still got a lot to offer and you can be sure that it will be a hot summer in the 3D-arena this year.

We can expect the first announcements of third-generation 3D-chips within the next couple of months and there's supposed to be a new quantum leap in 3D-performance and quality. Fill rates of 500-600 Mpixel/s, triangle rates of 20 million/s and more as well as the support of new quality enhancements will ask a lot more from game developers, CPU, memory, RAMDAC and even display technologies than we can imagine right now.

Let's focus on those chips, that provide today's top performance, TNT2, Voodoo3, Rage128, G400, PowerVR SG, Permedia3, etc. Voodoo3 was the first of those chips and the expectations were high when Voodoo3 was announced for the first time, since 3Dfx had the history of supplying two top performers in the past, the Voodoo in 1997 and the Voodoo2 in 1998. At Comdex 1998 the journalist and card-makers weren't too impressed with the Voodoo3-features though. Even the following marketing campaign wasn't able to convince more than a few fanatic 3Dfx followers, when 3Dfx threw in all of their arrogance and ignorance towards the criticism of 3D analysts and journalists about missing features of Voodoo3. What is sold as 'Voodoo3' today should rather carry the name 'Banshee2', since it's a straightforward development of Banshee with performance measures that are close to Voodoo2. We certainly don't want to forget improvements as digital flat panel support, higher resolutions and better RAMDAC and the price/performance ratio of Voodoo3 is also way better than Voodoo2 SLI. However, the absolute performance as well as the long list of missing features has disappointed a lot of hard core gamers.

The support of those two is of the highest importance for optimal 3D-game support. The majority of 1^st person 3D-action shooters is based on OpenGL, Quake2, Quake Arena, Unreal, Unreal Tournament, Halflife and many more are the typical examples. The majority of all other games is covered by DirectX, particularly flight simulators and racing games are mainly based on this 3D-interface from Microsoft. A 3D-card that doesn't support those two 3D-platforms is pretty much out of the question for any gamer.

The fill rates quoted by the chip vendors should be taken with a decent grain of salt. Firstly those data are 'synthetic', comparable to MIPS or MFLOPS of CPUs. The fill rate can be calculated from the 3D-chip clock, the number of independent pipelines of the chip, the graphics memory bandwidth and of course the hardwired features of the chip. The color depth, Z-buffer and particularly rendering quality varies significantly in between the vendors. A good example for this is TNT2, which has a lower theoretical fill rate than Voodoo3, but it scores higher frame rates in complex 3D scenes. Still the fill rate can give you some kind of idea of what kind of 3D-performance you can expect from a 3D-chip.

The onboard memory of each 3D-card has a very high impact on the overall 3D-performance. 16 MB is enough to run 3D applications at up to 1024x768 in 32 bit color depth (this is not valid for PVRSG, which needs less memory for that). Realistic 3D scenes at 1280x1024x32 bit color require 32 MB and that's what is state-of-the-art today. The memory clock still varies a lot between the different chips and so does the width of the memory interface, where currently 128 bit is state-of-the-art, moving towards 'dual-256 bit' = 2 interfaces with 128 bit or real 256 bit soon. The graphics memory bandwidth is an important limiting factor of the 3D-performance and you can say that a vendor who can safely handle highest memory clocks is well prepared for future developments.

More and more games are offering a true-color mode due to quality reasons, meaning that the displayed frames are using a 24 bit deep color depth (RGB, 16.7 million colors). This means that a huge amount of data needs to be computed for each frame, but the performance of the latest 3D-chips is starting to be capable of supplying this kind of computing power. A 3D-architecture that is limited to only 16 bit, as in case of the Voodoo3 requires a reduction of color information and is thus decreasing the quality and reality level. 32-bit color rendering with 32-bit Z-buffering will stay the optimal implementation for quite some time to come.

The RAMDAC is the link between the digital display information stored in the frame buffer and the analog CRT. The higher the bandwidth of the RAMDAC, the higher is the number of pixels that can be displayed on the CRT each second. This means that a faster RAMDAC is crucial for decent refresh rates, its quality is crucial for the quality of the displayed picture on the screen and only fast RAMDACs can supply really high resolutions of more than 2000x1500, which is most important for 2D applications. If the user prefers a flat panel screen, he should make sure that there is no digital-analog-converter (DAC) and then a analog-digital-converter (ADC) between the graphics memory (or frame buffer) and the flat panel, because it is not only a waste of silicon, but it also decreases the display quality by a serious amount. A flat panel requires digital data and should thus be supplied with nothing else than that. That requires a digital output of the graphics card, where DFP is one of the standards. For optimal 2D-quality you should therefore either look for a fast RAMDAC for a CRT or a digital output for a flat screen.

AGP vs. PCI is a question that doesn't need to be asked anymore today. AGP has succeeded in the upper performance area quite a while ago already. When AGP 4x will be released later on this year, we may finally see the visible advantage that many of us are still missing. A today's 3D-card should support AGP 2x to be at least able to handle the data bandwidth that's required for 'several million polygons/s'. Games with large textures are coming up, FlightSim 99 is only one example, so that the 3D-chip should also be able to do AGP-texturing, which is one of the things that Voodoo3 is not able to do.

So far about the requirements of a well performing 3D-chip or 3D-card. Of course there are a lot of other effects that 3D-chips can do today as well, like fog, transparency or different kind of shadings, but those 'common' features don't really need to be mentioned again. However, there are still cases where features are implemented at a low quality level, when 3D-chip makers rate frame rates over image quality.

Realism is what we really want, the users and game developers have to tell the 3D-chip makers what they expect. The hardware developers have made another leap forward in terms of realism with the latest generation of 3D-chips and 3D-games should follow soon to take advantage of it. A look at the CPU-roadmaps shows that we can expect a lot more performance in very short time.