There was the Albatron ATOP (never seen in the wild be me), which was the NVidia HSI bridge with a jumper configuration, achieving exactly that:

https://www.techpowerup.com/forums/threads/mo … atop-info.3272/

Here are some photos of the bracket being rotated, along with what I think is the riser I bought years ago. I don't know how it would function if the bracket doesn't fit.

Anyone willing to make this?

You would not gain much from using the AGP slot for that using a simple PCI to PCIe bridge.
It will omit all the AGP features (side-band addressing etc.) and you'll get AGP 1x speeds at best.

Also, I could swear I saw the same question + picture in another thread the other day.. But apparently this is your first post. That seems a bit fishy.

tehsiggi wrote on 2026-02-10, 10:08:

You would not gain much from using the AGP slot for that using a simple PCI to PCIe bridge.
It will omit all the AGP features (side-band addressing etc.) and you'll get AGP 1x speeds at best.

Also, I could swear I saw the same question + picture in another thread the other day.. But apparently this is your first post. That seems a bit fishy.

Well most PCI is only half of AGP 1x so PCIe on AGP would still be better than PCIe on PCI.

As for your suspicion about same question & picture, I agree. I think it was in the thread with title "PCIe card in AGP slot. Do such adapters exist?" but it looks like the comment has been deleted.

Glad I'm not making things up in my head. And yes, it was that thread.

Yeah, you'll get the benefit of 66MHz on AGP vs. 33MHz on PCI, but a bad bottleneck is a bad bottleneck.

The simplest way is AGP > PCI + PCI > PCIE.
Electrically it should work as expected but it will be clunky looking

myne wrote on 2026-02-11, 01:57:

The simplest way is AGP > PCI + PCI > PCIE.
Electrically it should work as expected but it will be clunky looking

I think level shifting (0.8V/1.5V->3.3V) would be required for working with modern AGP ports (4x or more).

RECNAS (AGP -> PCI) doesn't appear to have any level shifting logic. It mainly targeted older AGP (2x) boards which used 3.3V and no need to shift levels there.

The pci>pcie chip is presumably able to deal with voltages.
You'd have to check the datasheet, but I doubt they'd make one that couldn't be used on agp or pci

The agp>pci adapter would just be wiring

myne wrote on 2026-02-11, 01:57:

The simplest way is AGP > PCI + PCI > PCIE.
Electrically it should work as expected but it will be clunky looking

Deleted the old post so there wasn't a duplicate, this thread seemed more active. The pex 8114 supports up to 133mhz i believe so we could oc the pci/agp clock and run all pci slots with pex8112 china adapters. 33/66mhz limitation can be side stepped, and even a bottle necked 1080 ti will still do better than a 3850 agp surely, even without the agp specific features.

tehsiggi wrote on 2026-02-10, 10:08:

You would not gain much from using the AGP slot for that using a simple PCI to PCIe bridge.

If it works in a similar fashion that PCI-X-to-PCIe works, there will be massive increase in some configurations.
Re: A useless short analysis on PCIX to PCIe bridge performance on Tualatin
Re: A useless short analysis on PCIX to PCIe bridge performance on Tualatin

The Serpent Rider wrote on 2026-02-16, 08:44:

If it works in a similar fashion that PCI-X-to-PCIe works, there will be massive increase in some configurations. Re: A useless […]
Show full quote

tehsiggi wrote on 2026-02-10, 10:08:

You would not gain much from using the AGP slot for that using a simple PCI to PCIe bridge.

If it works in a similar fashion that PCI-X-to-PCIe works, there will be massive increase in some configurations.
Re: A useless short analysis on PCIX to PCIe bridge performance on Tualatin
Re: A useless short analysis on PCIX to PCIe bridge performance on Tualatin

Yes. doubling the clock will of course give more bandwidth.

Iirc we're speaking:

• PCI - 33MHz => ~ 133MB/s
• PCI (not X) - 66MHz => ~ 266MB/s
• AGP 1x => ~ 266MB/s
• AGP 2x => ~ 533MB/s
• AGP 4x => ~ 1.07 GB/s
• AGP 8x => ~ 2.13 GB/s

So you'll effectively (if your don't overclock the PCI bus) you'll be stuck at AGP 1x speeds. For overclocked PCI at 133MHz you'd get AGP 2x speeds.

That's still way slower than PCIe 1.0 x16.

I'd not be sure if it's worth that. You'd also have to get drivers etc. that work well with the cards (and their overhead) - It'd make for a funny experiment, but I don't think of much use beyond that.

People forget that the bus is mostly used in bursts if the card has enough ram.

It's a big burst to fill the ram during loading, and then a relative trickle of geometry and instructions during play.

Ie, the load times are likely to be awful, but the gameplay should be relatively decent assuming the cpu/driver can keep up.
It is likely to perform more consistently with a frame rate cap though.

myne wrote on 2026-02-18, 00:54:

People forget that the bus is mostly used in bursts if the card has enough ram. […]
Show full quote

People forget that the bus is mostly used in bursts if the card has enough ram.

It's a big burst to fill the ram during loading, and then a relative trickle of geometry and instructions during play.

Ie, the load times are likely to be awful, but the gameplay should be relatively decent assuming the cpu/driver can keep up.
It is likely to perform more consistently with a frame rate cap though.

I think this probably explained something with my current AIMB-865 setup, using a FirePro W4100 behind a PEX8112 on a PCI slot.
- Linux (EndeavourOS) works best with "amdgpu" driver, which has almost no issues even at 1080p 120Hz/144Hz resoultion. The default "radeon" driver behaves very buggy, however.
- Windows 7 works mostly fine on certain driver versions at 1080p 120Hz/144Hz resolution. No major issues apart from slowdowns when I drag a selection rectangle large enough.
- WEI on Win7 reports a score about 7.6 for Aero and Graphics using latest Radeon Enterprise drivers that work correctly (some versions appear to have known issues). With older Catalyst drivers, however, I get a score up to 6.8-6.9.

So PCI bus bottleneck does affect some functionality but not everything. Considering the high WEI score I'm getting on Win7, the bottleneck likely did not impair functions that actually matter.

I could not get that card working with WinXP due to Code 10 on that board, though. The card, which uses a smaller factor but otherwise identical to FirePro W600 GPU-wise, is not officially supported on XP anyway.

tehsiggi wrote on 2026-02-16, 19:40:

Yes. doubling the clock will of course give more bandwidth.

Well, it's a bit more complicated than that. From my experience, regular PCI suffers from all kinds of weird slowdowns which are not present on AGP even without any GART driver installed (i.e. simple PCI 66 mode). Like Quake 3 barely reaching 60 fps on any settings on variety of PCI cards. So benefits can far exceed 100% boost. I suspect that it has something to do with AGP (and PCI-X in some configurations) being a dedicated data lane on the north bridge for one device.