Performace of AGP vs PCI

[Bstansbury]

I posted a note the other day but the scope of my question was too broad.

Here is the key issue. On an good PC system motherbaord you have both PCI buss and a AGP port which is primarily to support high res. graphics adapters. However, the new ATI Radeon cards give you the choice of either.

The goal of all Video Capture cards is to capture video data and get it to the hard disk. For a PCI card, video data comes into the real of the card and is processed thru on card logic to get it to the display monitor, which in the case of ATI, is on the same card. So the display data never has to exit the card (unless it has to go to main memory for some reason). The captured video data has to go through a different path, off the card through either the PCI buss or the AGP buss to the motherboard and into main memory. One of these busses is faster than the other.

Once in memory the data will most likely be DMA'ed out to the ATA 66 or ATA 100 MBs EIDE buss to the hard disk.

The question is, since video capture has an extremely high demand on system performance, which adapter card would give the system better performance. Since in both adapters, the captured video data has to get onto the motherboard and into the main memory, which has the higher bandwidth into memory. We know the PCI buss is about 33 MBs. What is the Bandwidth of the AGP buss.

Thanks to anyone willing to add factural information.

[Hiei]

Agp bus is faster. How much faster is AGP than PCI? A 33-MHz 32-bit PCI bus supports up to 132-MBps throughput. AGP is also a 32-bit design, but it runs at speeds up to 133 MHz--four times as fast--so it has a maximum transfer rate of 532 MBps. (This is still twice as fast as the rate of a 66-MHz 32-bit PCI bus.) Best of all, however, is that the graphics card on the AGP bus does not have to compete with any other devices to get access to its data.

[JHebert]

While AGP is a much faster bus than PCI, usually the bottleneck in video capture performance is not the PCI or AGP bus speed, but is the transfer rate of your hard drive(s).
Uncompressed video at 640x480 and 29.97 fps requires about 28MB/s (RGB) or 18MB/s (YUY2) which is well within PCI transfer capability of 132MB/s.
However, even a fast single hard drive can struggle to maintain these sustained transfer rates for the duration of a video capture without dropping frames. Compressing video with video codecs reduces the requirements for drive speed and storage space but still the hard drive is usually the bottleneck. Using a RAID system of multiple striped hard drives can give you much faster disk performance for optimal video capture and editing performance.
When choosing a capture card I would focus more on the chipset and other components and the quality of drivers and software instead of AGP vs PCI bus. However for graphics performance in 3D gaming etc AGP has a definite advantage over PCI so if you're getting a combination graphics card with TV-in/out like an All-in-Wonder I would get an AGP version over PCI.

[Ronald_B]

<TABLE BORDER=0 ALIGN=CENTER WIDTH=85%><TR><TD><font size=-1>Quote:</font><HR size=1 color=black></TD></TR><TR><TD><FONT SIZE=-1><BLOCKQUOTE>
On 2002-01-14 08:03:08, Hiei wrote:
Agp bus is faster. How much faster is AGP than PCI? A 33-MHz 32-bit PCI bus supports up to 132-MBps throughput. AGP is also a 32-bit design, but it runs at speeds up to 133 MHz--four times as fast--so it has a maximum transfer rate of 532 MBps. (This is still twice as fast as the rate of a 66-MHz 32-bit PCI bus.) Best of all, however, is that the graphics card on the AGP bus does not have to compete with any other devices to get access to its data.

</BLOCKQUOTE></FONT></TD></TR><TR><TD><HR size=1 color=black></TD></TR></TABLE>

Correct, that's why there are wait states on the PCI bus.

[Bstansbury]

Thanks for all the answers. It seem clear the AGP is the best solution.

I did see the motherboard layout in my ABIT manual and the bottle neck seems to be the Via KT133 chip. All busses go through that chip before they can get to memory,..but in the case of the PCI buss, it has to share the PCI buss for data going into the memory and also back out of memory, across the PCI(internal) buss out to the IDE ports. This is double duty for that path.

Using AGP, the path in goes through the KT 133 chip into memory and back out of memory through the KT 133 again and then across the internal PCI buss to the IDE controller and finally down to the disk drives. The key is that name of the chip is KT 133 meaning that is it's clock rate.

That kind of brings up another question,... If the clock rate of the chip is 133 MHz, and both the PCI buss and teh AGP buss are 32 bits wide, how is one faster than the other.

Also, what does AGP 1x/2x/4x mean???

Thanks again for you answers. I welcome any other comments and I'm always trying to get educated about this stuff. When I worked at IBM, there was a saying, "There is no end to education". How true that is.

[JHebert]

AGP bus runs at 66Mhz on a 32 bit wide bus. AGP2X is "double pumped" meaning it transmits data on the rising and falling clock cycle for effectively double the bandwidth for the same bus speed. AGP 4X is "quad-pumped" for four times the bandwidth. The maximum data transfer rates are as follows:
AGP 1X = 266MB/sec
AGP 2X = 533MB/sec
AGP 4X = 1.066GB/sec

PCI bus is normally 33Mhz on a 32 bit wide bus. There are also PCI versions at 66Mhz on a 32 bit bus, 33Mhz on a 64 bit bus, and 66Mhz on a 64 bit bus. These PCI variants are typically found on servers and high end workstation systems. The maximum data transfer rates are as follows:
PCI 33Mhz / 32 bit = 133MB/sec
PCI 66Mhz / 32 bit = 266MB/sec
PCI 33Mhz / 64 bit = 266MB/sec
PCI 66Mhz / 64 bit = 512MB/sec

The speed of 133Mhz for the KT133 chipset is the memory bus speed, which is the interface between the KT133 and main memory. Newer chipsets like KT266 support DDR memory for an effective 266Mhz memory bus speed.
The KT133 has a 200Mhz frontside bus speed, which is the interface between the KT133 and the Athlon CPU. Actually its really 100Mhz but is double-pumped similar to how AGP works, for double the effective bandwidth. The newer KT266 supports 133Mhz (266Mhz effective) frontside bus speed for Athlon.
The Pentium 4 has a "quad pumped" frontside bus which runs at 100Mhz x 4 for effective 400Mhz speed. The older Pentium 3 are limited to 100Mhz or 133Mhz frontside bus and the older Celeron was only 66Mhz.

There are bus speed multipliers in the KT133 chipset that run the PCI and AGP buses as a fraction of the frontside bus speed. For example for a FSB of 133Mhz the AGP bus runs at a 1/2 the speed and the PCI bus runs at 1/4 the speed. These multipliers change if you run the FSB at 66Mhz or 100Mhz to keep the AGP and PCI bus at the correct speeds.
Problems can arise if you try to overclock your system by increasing the FSB with non-standard settings like 112Mhz or 147Mhz. You end up with AGP and PCI buses running at overclocked speeds because they are tied to a limited choice of bus multipliers. Some AGP video cards and many PCI cards do not tolerate running at overclocked settings and cause stability problems.
When you overclock by changing the CPU multiplier instead of the FSB speed you avoid these problems because the AGP and PCI bus speeds are not affected. This is only possible on AMD CPUs because Intel CPUs are "multiplier locked" and do not allow this type of overclocking.

[Bstansbury]

Thanks for the great answers. I bet you are a motherboard designer, or at least one of the chipset designer. This is the kind of factural data that can really provide clear and understandable answers for all of us.

Since you've seen the discussion above... what is your view of the Buss conflicts since all of these busses have to go through one chip,..i.e., KT133. Even at moderate data rates in and out of main memory, will there be enough burst conflicts to cause dropping of frames.

Also, if AGP 4x video data comes into main memory (via KT133) and at the same time captured data has to exit main memory out through KT133 to the PCI buss, are there buffers in the KT 133 to keep the data moving on both busses without overrun.

I appreciate the time you've taken to answer these questions and I hope that many people get to benefit from them. Thanks again.