MattRocks wrote on Today, 20:11:

Grem Five wrote on Today, 20:04:

I'm confused with this 3.3 volt talk, the board PC@LIVE posted the manual states "ASRock Graphics Interface Slot (1.5V_AGP1) "

I don't know. I steered away from voltages - just saying the plastic slot itself should physically prevent users inserting cards that are electrically incompatible. I'm trying to focus on PCI/AGP bandwidth, when it matters, and when it doesn't.

I dont know I was just confused because in some posts there was being thrown out 3.3 volts and 33 mzh but being this board mentions 1.5 volt slot and if this AGI slot is based on the pci slots on the board which are spec 2.2 which can run at 66 mzh. <shrug> that sounds like at least AGP 4x and from what I remember back in the day most people couldnt notice a difference between 4x and 8x.

Edit: I took look at the TRW page for this board

Known issues […]
Show full quote

Known issues

Fake AGP slot

Any AGP cards inserted into this slot operate in PCI 33Mhz Fallback mode.

DO NOT use 1.5V exclusive AGP cards (rear notch away from IO side) in this board.
ONLY use true universal AGP cards with both slots, these support 1.5V and 3.3V operation.
This is a PCI bus connected to an AGP slot.
The PCI bus operates at 3.3V and may damage 1.5V exclusive cards.
AGP cards are backwards compatible with PCI signalling logic.
Performance is greatly reduced vs a true AGP bus.

Those know issues listed dont seem to be anything I gleam from looking at the manual but they seem to contradict what they decided to put in manual, I can see the voltage part but falling back to 33 Mzh doesnt really make sense unless a slower pci card is put in one of the other pci slots that is only a 33 mzh card. I mean even the agp slot is keyed for 1.5 volts. Guess ASRock was a bit shady back then as well.

I'm reading it as PCI 133 MB/s only because for a similar board Retroweb says, "Any AGP cards inserted into this slot operate in PCI 33Mhz Fallback mode"
https://theretroweb.com/motherboards/s/asrock-p4dual-915gl

But, Retroweb is not the authority and there is no official Intel AGP behaviour called "PCI 33Mhz Fallback mode". Intel says it is always 66 MHz but that the messages transmitted at 66 MHz can be with or without the AGP extensions to PCI,

"The A.G.P. interface specification ... provides three significant performance extensions or enhancements to which are intended to the PCI specification ... These enhancements are realized through the use of “sideband” signals."

So, if you disable sideband, you disable the PCI extensions! Typically, at least in AGP 1.0 era, we mostly disabled the sideband. Furthermore, Intel specifications show is that PCI signals and transactions are within in AGP standards,

"Figure 3-30 is a basic PCI transaction on the A.G.P. interface. If the PCI agent is a non A.G.P. compliant master, it ignores the ST[2::0] signals and everything appears to be a PCI bus."

"The earliest the A.G.P. compliant master can request a PCI transaction is on clock 3, and the earliest the arbiter can grant permission is on clock 4, which allows the PCI compliant master to initiate its request on clock 6. The two clocks between the PCI transaction and the read data is caused because of potential contention on TRDY#. This can occur when the PCI compliant master is the core logic and the target is the A.G.P. compliant master."
https://www.dcs.ed.ac.uk/home/ecole/agp/agp10.pdf

So, the Intel AGP 1.0 spec implies that all AGP cards must be aware of PCI-style signalling and those of us old enough to have used AGP 1.0 were disabling the AGP-specific extensions to achieve stability.

AGI seems to exploit the old PCI-compatible subset of AGP, but that does not mean later AGP 3.0 cards/drivers were optimised to exercise the PCI-style transactions.

Now it seems the 33 MHz detail might be erroneous. It may be that when the AGP cards try to use PCI extensions on the AGI slot to make AGP transactions, the AGP card needs to gracefully fallback to making PCI transactions instead - exactly as the AGP card would have to do if you entered BIOS setup and disabled AGP sideband, disabled AGP aperture, disable AGP fast writes, etc.

MattRocks wrote on Today, 20:03:

Socket3 wrote on Today, 19:53:

133MHz is a huge bottleneck for DX7+ graphics cards. Take the FX 5200 for example. I have two almost identical cards - a Zotac P […]
Show full quote

MattRocks wrote on Today, 18:10:

Ok, but AGP 1.0 was 3.3V too. Is there any objection to treating AGI vs AGP as PCI vs AGP?

My view is that a technical comparison should boil down to 8xAGP 2133 MB/s vs PCI 133 MB/s, and to have any real-world differentiator there needs to be evidence that 133 MB/s is a bottleneck during DX9 gameplay. I have never seen any evidence of that. My challenge is to surface that evidence, which might exist! I only know that during the DX6 generation the 133 MB/s limit was more imagined bottleneck than actual bottleneck.

Radeon 9700 Pro pumps data with 19.4 GB/s VRAM bandwidth, and I suspect games running on a 9700 don't give one hoot about the bus bandwidth. And, I say that because games running on a Voodoo3 with 2.3 GB/s VRAM bandwidth didn't give one hoot about bus bandwidth.

133MHz is a huge bottleneck for DX7+ graphics cards. Take the FX 5200 for example. I have two almost identical cards - a Zotac PCI FX 5200 (blue PCB) - 256MB of 5ns DDR over a 128 bit bus, running at 333Mhz vram and 250MHz core - and an Inno3D FX5200 "ultimate" (that's the name of the card, it's on the box) - black PCB, same exact elixir 5ns memory with the exact same layout (all 8 vram chips on the front of the card) only AGP. The inno3d card is clocked at 250mhz core and 400Mhz vram. On my Pentium 4 3400Mhz testbench (abit AS8 mainboard) the Zotac PCI card scores a whopping ~5900pts in 3D Mark 01, and that's after I overclocked the vram to 400Mhz to match the inno3d card (it's the same ram so why not give the PCI card a fair shot). The AGP Inno3D at the same clocks scored 8300 pts - 30% difference. If that's not a bottleneck, I don't know what is.

Furthermore, in UT2004 the Inno3D AGP card offers somewhat playable performance @ 1024x768 medium settings - the PCI card is unfortunately stutterfest central. Similar FPS but horrible frametimes making the game unplayable.

I'll post proof in the weekend. I have the benchmarks saved. I don't know if vogons will let me upload video, but I'll film both cards in UT2004 and try to post it. I'll make a new thread in the "video" section.

" Zotac PCI card scores a whopping ~5900pts in 3D Mark 01 ... AGP Inno3D at the same clocks scored 8300 pts"

In other words, the slower PCI bus bandwidth is not the limiting factor? You appear to be strengthening my argument by saying a PCI card outpaces the AGP card

What now? The Zotac is PCI, and is 30% slower than the Inno3D AGP card. Please read again.

MattRocks wrote on Today, 20:03:

- exactly the same thing we saw with Voodoo3 PCI cards outpacing Voodoo3 AGP cards.

From my testing on a fast PC (KT333 / Athlon XP 2600+) the PCI version looses to the AGP Voodoo 3 3000. On a period correct machine there is no difference.

Socket3 wrote on Today, 21:41:

MattRocks wrote on Today, 20:03:

Socket3 wrote on Today, 19:53:

133MHz is a huge bottleneck for DX7+ graphics cards. Take the FX 5200 for example. I have two almost identical cards - a Zotac PCI FX 5200 (blue PCB) - 256MB of 5ns DDR over a 128 bit bus, running at 333Mhz vram and 250MHz core - and an Inno3D FX5200 "ultimate" (that's the name of the card, it's on the box) - black PCB, same exact elixir 5ns memory with the exact same layout (all 8 vram chips on the front of the card) only AGP. The inno3d card is clocked at 250mhz core and 400Mhz vram. On my Pentium 4 3400Mhz testbench (abit AS8 mainboard) the Zotac PCI card scores a whopping ~5900pts in 3D Mark 01, and that's after I overclocked the vram to 400Mhz to match the inno3d card (it's the same ram so why not give the PCI card a fair shot). The AGP Inno3D at the same clocks scored 8300 pts - 30% difference. If that's not a bottleneck, I don't know what is.

Furthermore, in UT2004 the Inno3D AGP card offers somewhat playable performance @ 1024x768 medium settings - the PCI card is unfortunately stutterfest central. Similar FPS but horrible frametimes making the game unplayable.

I'll post proof in the weekend. I have the benchmarks saved. I don't know if vogons will let me upload video, but I'll film both cards in UT2004 and try to post it. I'll make a new thread in the "video" section.

" Zotac PCI card scores a whopping ~5900pts in 3D Mark 01 ... AGP Inno3D at the same clocks scored 8300 pts"

In other words, the slower PCI bus bandwidth is not the limiting factor? You appear to be strengthening my argument by saying a PCI card outpaces the AGP card - exactly the same thing we saw with Voodoo3 PCI cards outpacing Voodoo3 AGP cards.

What now? The Zotac is PCI, and is 30% slower than the Inno3D AGP card.

I explained on the previous page that I made a numeric dyslexic misread. Shit happens.

MattRocks wrote on Today, 21:43:

Socket3 wrote on Today, 21:41:

MattRocks wrote on Today, 20:03:

" Zotac PCI card scores a whopping ~5900pts in 3D Mark 01 ... AGP Inno3D at the same clocks scored 8300 pts"

In other words, the slower PCI bus bandwidth is not the limiting factor? You appear to be strengthening my argument by saying a PCI card outpaces the AGP card - exactly the same thing we saw with Voodoo3 PCI cards outpacing Voodoo3 AGP cards.

What now? The Zotac is PCI, and is 30% slower than the Inno3D AGP card.

I explained on the previous page that I made a numeric dyslexic misread. Shit happens.

oh, sorry about that.