AGP vs PCI

[bluedanube]

1. Is there a PCI card that is considered almost as good as a agp card?

2. I own a dell optiplex which only has pci extension slots. Can I add an agp card, somehow? :(

[NDD]

1. Yes. Alsmost all PCI cards :)
AGP doesn't have so much to do with graphics/3D speed as some/most tend to think. AGP texturing, one of the greatest things AGP was developed for, lost its advantage when more and more videocards came equiped with more memory, 64MB+
2. No. Some PCI-to-AGP convertors DO exist, but I seriously doubt they're good :rolleyes:

[KraZy_SkitZy]

My old TNT2 32mb PCI is as good as any other TNT2 32mb AGP 4X..... In my opinion... ;)

[Peter M]

PCI slots have a much lower power budget than AGP 2.0 slots - this is what limits PCI card performance nowadays.

Then, the PCI bus is much further aback in bus priorities. AGP hangs directly off the North Bridge, thus has a very short and fast path to CPU and main memory. PCI traffic has to go from North to South and then out to PCI, which is a long way.

Finally, PCI graphics cards are often observed to eat way too much bandwidth on the north/south interconnect and PCI, which does things like throw audio replay off rail, make ethernet lose connection and such.

regards, Peter

[causticVapor]

Hmm.. interesting. I always thought PCI graphics cards, being the bandwidth monsters they are, would hog the bus and make sound cards' bursts to main memory interrupted... as for power, they could have always provided a +5v connector on the card if that was an issue...

The card companies probably thought it would be worthless to port really high-performance cards to the PCI bus, because first of all, they would suck up the bandwidth with all their streaming (64 clocks or so??) and be bottlenecked anyhow.

Wait a second, don't PCI graphics cards limit their bus mastering to the latency timer set in the BIOS? I've heard that a value of under 100 is detrimental to a modern graphic's card's performance but above 32 or so, the card is almost stealing the entire bus for the majority of the time and controller cards, nics, sound cards, and all the others would not get a chance at all.

Yep... it's not so much the 133MB/s limitation as it is the "competition" to the southbridge (nb on older chipsets)....

Here's a question - if you only had the graphics card on the PCI bus with no other devices, then could you garner near-AGP 0.5x performance by setting the latency (bus mastering value) very high? Of course, it would be limited to 133MB/s and then whatever the vlink's bandwidth to the northbridge is, or on older boards, simply 133MB/s to the northbridge.

According to some sites, however, latencies of 248 clocks or so can impose a performance penalty :confused: even though that is a default for the nvidia cards :confused: This, again is AGP, so I may be going way off topic....

[tony_j15]

AGP runs at 1gig a second. PCI runs at 133MB a second. Figure out which is better.:rolleyes:

[causticVapor]

No, it's not exactly that, sir -- and I'm not comparing PCI to AGP to see which is "better."

Only AGP 4x transfers at 1GB/s.

[Peter M]

Originally posted by tony_j15
AGP runs at 1gig a second. PCI runs at 133MB a second. Figure out which is better.:rolleyes:

You should first figure out that AGP transfers hardly ever get used. Most stuff moving to an AGP graphics card while a scene is rendered is geometry data, these don't use AGP transfers. So the raw speed gain is just like from 33 to 66 MHz PCI. It's the more direct attachment to the system core that makes AGP actually faster than PCI.

As for the latency timer, first of all they're a niceness feature. The bus owner doesn't HAVE to yield after that time elapsed, it SHOULD. If the card feels there's an urgent need to transfer more, it can do so.
Side note: Intel 8xx and VIA chipsets can actively pull bus ownership from any card at any time and re-arbitrate. This improves fairness, but worsens throughput. VIA went with that option for a long time, only to reverse when benchmark fetishists started shouting BAD THROUGHPUT ON VIA CHIPSETS.

It's a saturation problem - if the actual bandwidth demand approaches the overall ceiling, there's little you can do. We have the (undecoded) DVD data moving from IDE to system RAM, then back to the GPU and/or CPU for decoding, and then finally into graphics card overlay buffer, plus the sound data. That makes for quite a lot of data, and it's all data that need to travel halfway isochronously to pleasure the viewer.

All this is quite manageable as long as the graphics card does the decoding or the CPU and RAM are halfway fast, and is on a bus that isn't at some point concurrent with the sound and IDE datapath. And that's where the strength of AGP lies here - it has its own north bridge connection, while the PCI data travel the same inter-chipset bus as the IDE and sound data do.

regards, Peter

[Beeblequix]

wow Mr. Missel, that's some impressive explanation. Have a pastry *ß.Q. hands a raspberry filled danish to Peter*
:t

[NDD]

With Peter, you can't expect less :)

[Giblet Plus!]

Originally posted by causticVapor
as for power, they could have always provided a +5v connector on the card if that was an issue...


Voodoo 5, anyone?

[causticVapor]

Well, yes :D

That card would have some serious starvation problems otherwise

Power monger :x