During Computex we saw photos of several Intel 925x boards equipped with both PCI Express and AGP slots.  Since Intel?s new chipset doesn?t offer simultaneous support for PCI-E and AGP, there was a serious question as to how Gigabyte and ECS (the two manufacturers with boards on display) were accomplishing it. 

At the time, I speculated to Tuan that one or both of the companies might have used an AGP connector on a PCI slot, as the only other solution would seem to be an AGP-to-PCI-Express bridge chip that left only one of the slots useable at a time, the latter option being only a proposed theory, as no such design exists to our knowledge).

As Anand revealed here, both manufacturers have gone the PCI route, and hung an AGP port off the PCI bus.  This is perfectly feasible from a manufacturing standpoint, but the performance penalty should be huge.  Amazingly, SiS claimed only a 5-20% performance penalty from using this solution.

The PCI bus provides 133 MB/s of bandwidth shared between all the PCI slots.  AGP 8x provides a dedicated 2.1 GB/s.  For ECS to claim that a 93.7% drop in available bandwidth only equated to a 20% drop in performance indicated one of two things:  Either the usefulness and necessity of AGP had been fundamentally misrepresented through the years, or ECS was ludicrously optimistic when estimating the performance penalty of an AGP-compatible PCI solution. 

My curiosity piqued, I decided to investigate just how much AGP bandwidth affected modern game and application performance, and, by extension, how much PCI-Express is likely to improve it.  While we won?t be able to directly predict PCI-E performance based on AGP scaling, we can use one to establish a rough trend for the other.

Before we delve into a discussion of AGP performance lets discuss briefly what it is.

The History of AGP:

AGP was developed in the mid-to-late 1990s to address the twin problems of limited PCI bandwidth and the very high cost of video RAM.  At the time, 8 meg of VRAM was the practical limit that could be built into the PCB.  If a card used a 4 meg frame buffer, that left only 4 meg for any sort of processing work.  This was sufficient for basic 2D, but not for higher-level 3D applications. 

AGP addressed both issues simultaneously by creating a dedicated port between video card and processor, a port whose bandwidth would be reserved for only the video card.  In order to facilitate the usage of system memory for video performance, AGP was designed to use DMA.  Although we tend to view ?integrated? video systems with a jaundiced eye today, the original AGP standard was explicitly designed with them in mind.  Given the primacy of 2D applications and the limited amount of available VRAM when AGP was designed, it was believed that frame buffer memory would remain on the video card, while AGP facilitated the use of system memory for 3D performance. 

The original AGP specification was based on the PCI 2.1 spec and ran on a 66 MHz bus.  This explains why its possible for an AGP slot to be fitted on a standard PCI system without much effort?fundamentally, the bus architecture is identical.

The original AGP 1x specification ran at 66 MHz and provided 266 MB/s of bandwidth.  AGP 2x offered 533 MB/s of bandwidth and first appeared in August 1997 with the release of Intel?s 440LX chipset.  AGP 4x first appeared in late 1999 but the high price of RDRAM and unattractiveness of i820 combined with the low performance of VIA?s AGP solutions probably slowed adoption of the standard.  By the summer of 2000 and the launch of Intel?s i815, however, a wide variety of mature AGP 4x chipsets were available.

In August of 2002 VIA launched their KT400 with AGP 8x support.  In five years, available AGP bandwidth had quadrupled from 533 MB/s on the 440LX to 2.1 GB/s using VIA?s KT400.  In a few weeks, Intel will launch the first PCI-Express motherboards whose x16 PCI-E links will offer 4.1 GB/s of memory bandwidth.  In seven years we?ve nearly octupled total video card bandwidth?but how necessary is the gain?  Is PCI-Express needed in the video card industry? 

Analysis, Conclusions.

Analysis:  AGP on PCI Performance Penalty Likely Mis-Represented:

We don?t have the whole story on how the new boards from ECS or Gigabyte are designed, so I wouldn?t draw firm conclusions.  There are, however, several reasons to think that ECS? performance claims are wrong and drawn from a very limited sample.

A true AGP-on-PCI solution is going to lack a GART driver and be limited to only 33 MHz of shared-bus PCI.  In the majority of our tests moving from AGP .67 to AGP 1x and installing GART made the most difference out of any change.  Any AGP-on-PCI solution on a standard PCI slot is going to be even more bandwidth limited (.5x AGP vs. our .67x) and hence, crippled from the start.  The loss of GART is an additional performance whack that?ll hit performance far harder than 20%.

Even assuming motherboard manufacturers used an expensive, PCI-X solution, there?s still the problem of no GART and the fact that PCI-X is still a shared bus implementation.  Using ANY 33-MHz PCI card on such a board would promptly prevent the PCI-X bus from running at its standard 66 MHz. A 33 MHz, 64-bit PCI-X-AGP card would have roughly AGP 1x bandwidth, and thus most games would be playable--but a considerable number of titles want 2x at a minimum to really see performance--and of course, there's the loss of GART.

Given the bandwidth chokepoint and lack of drivers we?d expect any AGP-on-standard-PCI solution to cut performance 50-70% in almost any game or professional application.  Even UT2K3 *will* throttle downwards if AGP bandwidth is cut severely enough.  AGP on PCI-X would offer more bandwidth and higher raw performance, but also packs its own penalties. 

Bottom line?  You?re probably going to be better off staying away.  Far away.  Very far.

Conclusion:  PCI-Express on Desktop Likely to Offer Very Little Performance:

Is PCI-Express a bad idea?  Not at all.  It cuts manufacturing costs and simplifies trace design; from that perspective alone it?s a smart advance.  But don?t confuse ?smart advance? with ?performance advantage.?  Based on the numbers we?ve seen here, the average game and application (even the modern ones) typically do not take advantage of AGP 8x.  None of our games showed more than a small increase.  Professional applications are a different story, there we saw boosts of 30-40%.  Its also been suggested that certain specified applications, like video editing, might run more smoothly over the PCI-Express system.

For the gamer, however, PCI-Express is going to be a non-issue.  Given that most games today barely stress the AGP 2x bus, moving to double AGP 8x isn?t going to do anything.  It?s the equivalent of widening a highway no one drives on.  Had AGP developed in the direction it was intended to things might be different, but it didn?t.  Now, to some extent, it?s a solution to a problem no one has. 

If you?re still running on an original Athlon, P3, or Socket 423 P4, than a PCI-Express system might not be a bad idea, but if you?ve got something newer and aren?t using a specialized application demonstrated to gain from PCI-E, we wouldn?t bother planning to upgrade.  You?ll just be throwing money away after hype.  The reality is, in many cases we're barely pushing AGP 2x, sometimes 4x.  8x, despite now being two years old, is mostly hot air.

< previous - (6 of 6)

---

• System COnfiguration--Slowing Down to AGP 1x. (1 of 6)
• System Configuration and Details, Cont. (2 of 6)
• 3DMark 2001, 2003, AquaMark, Quake 3. (3 of 6)
• Comanche4, GunMetal, Final Fantasy, SPEC. (4 of 6)
• Unreal Tournament 2003 & 2004, Far Cry, Painkiller. (5 of 6)
• Analysis, Conclusions. (6 of 6)