I ran some tests with Quake 1 on a fast Pentium D 3.4Ghz, to rule out any CPU bottlenecks. There was some kind of incompatibility with the GeForce FX series cards (PCI/PCIe), but performance was consistent when tested with a Voodoo 4 4500 PCI and a Radeon X300 PCIe.

The test results were as follows:

66 Mhz - 29.9, 30.8
75 Mhz - 32.6, 33.6
80 Mhz - 35.9, 36.0
90 Mhz - 39.3, 39.6

These tests were run at 640x480x16. Performance was consistently higher than with 32bit.

Performance seems to be pretty linear with clock speed for this test. This is very promising. I don't have any faster clock crystals to test with beyond 90 Mhz.

I didn't experience any graphical corruption or anomalies at any clock speed.

Just ordered one of those PCI ribbon cable monstrosities. Going to finally try this, these results are amazing

For reference, this is the ribbon that maxtherabbit is referring to.

PowerVR PCX2 piggyback

Kahenraz wrote on 2022-06-06, 00:22:

I ran some tests with Quake 1 on a fast Pentium D 3.4Ghz, to rule out any CPU bottlenecks. There was some kind of incompatibilit […]
Show full quote

I ran some tests with Quake 1 on a fast Pentium D 3.4Ghz, to rule out any CPU bottlenecks. There was some kind of incompatibility with the GeForce FX series cards (PCI/PCIe), but performance was consistent when tested with a Voodoo 4 4500 PCI and a Radeon X300 PCIe.

The test results were as follows:

66 Mhz - 29.9, 30.8
75 Mhz - 32.6, 33.6
80 Mhz - 35.9, 36.0
90 Mhz - 39.3, 39.6

These tests were run at 640x480x16. Performance was consistently higher than with 32bit.

Performance seems to be pretty linear with clock speed for this test. This is very promising. I don't have any faster clock crystals to test with beyond 90 Mhz.

I didn't experience any graphical corruption or anomalies at any clock speed.

Screenshot_20220605-202734_Chrome.jpg

Thank you! Very interesting

Kahenraz wrote on 2022-06-04, 19:18:

Tested working at up to 90 Mhz. This larger heatsink keeps the chip quote cool. […]
Show full quote

Tested working at up to 90 Mhz. This larger heatsink keeps the chip quote cool.

I haven't heard anyone overclocking a PCX2 this high. It works great.

20220604_151331.jpg

20220604_151439.jpg

I struggle to find a 90Mhz oscillator. Where did you find it ?
I'd be interested by the socket's reference too. For now I couldn't find one as modular as yours.

I got ordered this crystal from a seller on AliExpress that specializes in them. I'll see if I can find it for you. At the time, I wanted to order other speeds, but this was during the height of the pandemic and he didn't know when they would be back in stock.

This particular layout is by VideoLogic. The other layout is by Matrox. There are different revisions of the VideoLogic PCB, with some slight component variation, but they are all pretty much the same.

What is different about yours? Can you share a photo?

Kahenraz wrote on 2022-06-20, 18:09:

I got ordered this crystal from a seller on AliExpress that specializes in them. I'll see if I can find it for you. At the time, I wanted to order other speeds, but this was during the height of the pandemic and he didn't know when they would be back in stock.

This particular layout is by VideoLogic. The other layout is by Matrox. There are different revisions of the VideoLogic PCB, with some slight component variation, but they are all pretty much the same.

What is different about yours? Can you share a photo?

Thanks for your reply.
This is this one Saving a PowerVR card
I meant your socket with multiple choices, it's really convenient. I admit I didn't search for very long but I know nothing about electronics (though I know how to solder/desolder).

These are solderable pin headers. I removed the old crystal and added the pin headers so that I could easily swap in new crystals for testing.

Here are some photos of the part. I've also included an eBay listing, but you may be able to find it cheaper elsewhere. I just did a quick search for you.

https://www.ebay.com/itm/373833415253

Hi guys !

Thanks Kahrenraz's explanations, I'm now running my Videologic Apocalypse 3Dx at 90MHz !
For cooling, i bought this radiator.
I removed the provided tape. Instead, I used thermal glue (two very small dots on two edges) and AC MX-5 thermal paste (it's pretty adhesive by itself).
Then I added a fan of the rad's size. This fan had to be silenced so I use a fan silencer from Noctua. It's directly on the Powerleap adaptor, because I was the closest source.
On top of both primary cards, two Noctua fans at full speed (it's silent anyways).

The noisiest parts of the PC are by far the power supply and the HDD.

Cooling seems to be adequate : the PVR card's radiator is only 5°C warmer than the primary card's when benchmarking Unreal since 20 minutes (the primary card is also getting air from the top Noctua fans).
It's 35C. For the record, on the chip's surface, it's 50°C hot at stock speed. Of course, a radiator's temp isn't a DIE's... But it's encouraging.

I didn't seriously benchmark it for now as I need to find the best ATi drivers (ATi cards are the fastest cards I've seen with PowerVR's, with excellent compatibility, but it depends a lot on drivers), but the gap is easily noticeable, as expected.

I'll make some benchmarks and captures soon.

Kahenraz wrote on 2022-06-06, 00:22:

I ran some tests with Quake 1 on a fast Pentium D 3.4Ghz, to rule out any CPU bottlenecks. There was some kind of incompatibilit […]
Show full quote

I ran some tests with Quake 1 on a fast Pentium D 3.4Ghz, to rule out any CPU bottlenecks. There was some kind of incompatibility with the GeForce FX series cards (PCI/PCIe), but performance was consistent when tested with a Voodoo 4 4500 PCI and a Radeon X300 PCIe.

The test results were as follows:

66 Mhz - 29.9, 30.8
75 Mhz - 32.6, 33.6
80 Mhz - 35.9, 36.0
90 Mhz - 39.3, 39.6

These tests were run at 640x480x16. Performance was consistently higher than with 32bit.

Performance seems to be pretty linear with clock speed for this test. This is very promising. I don't have any faster clock crystals to test with beyond 90 Mhz.

I didn't experience any graphical corruption or anomalies at any clock speed.

Screenshot_20220605-202734_Chrome.jpg

Here are my results with a Pentium Tualatin 1266 @1453MHz (153MHz FSB) :
Demo1 : 36.5
Demo2 : 36.6
Demo3 : 35.4

It seems that the difference between my CPU and yours is like overclocking the PCX2 by 10MHz. I have similar performance to yours when it was running @80MHz. I'd say it "scales", even if the difference of performance between our CPUs is huge.

Of course I made the benchmarks in 640x480 16bit colors like you. Which demo did you run ?

It would be interesting to determine at what speed the CPU becomes the bottleneck on these overclocks. How low of a spec CPU can I use and still maintain these numbers?

I'm making a Tillamook based PC which will have a PowerVR PCX1. I can make some tests with my overclocked PCX2 from 200 too 400MHz and see how it performs.

JustJulião wrote on 2023-03-25, 10:58:

Of course I made the benchmarks in 640x480 16bit colors like you. Which demo did you run ?

This was a long time ago and I can't remember.

I finally got around to modifying my m3d. I desoldered resistor R1 (a 330R pulldown to ground on FS0) and soldered in its place a 330R to Vdd. This gives me 80MHz (FS0 and FS1 both high) per the AV7-05 PLL datasheet. Unfortunately I only benchmarked quake 2 before the mod and found absolutely no improvement. I was getting 15.5 FPS on demo1 640x480 on my K6-2 @375MHz using the 3dnow patch both before and after. Can anyone who has a similarly speced system share a quake 2 benchmark to see where my bottleneck is please?

(I didnt end up using the ribbon cable PCI extender thing, the card is just installed normally)

maxtherabbit wrote on 2023-11-26, 02:03:

I finally got around to modifying my m3d. I desoldered resistor R1 (a 330R pulldown to ground on FS0) and soldered in its place a 330R to Vdd. This gives me 80MHz (FS0 and FS1 both high) per the AV7-05 PLL datasheet. Unfortunately I only benchmarked quake 2 before the mod and found absolutely no improvement. I was getting 15.5 FPS on demo1 640x480 on my K6-2 @375MHz using the 3dnow patch both before and after. Can anyone who has a similarly speced system share a quake 2 benchmark to see where my bottleneck is please?

(I didnt end up using the ribbon cable PCI extender thing, the card is just installed normally)

I didn't even know that we could use the 3Dnow patch with a 3D accelerator. I systematically failed using both at the same time, to the point I convinced myself that it was meant for software mode.
Besides my 90MHz PCX2 I only have a PCX1 currently, I'll be able to tell you for a regular PCX2 in the next few weeks if nobody does before.
With sound ON ? Which video options ? Which chipset ? Which FSB speed ? What amount of L3 cache ?

Sound on, 430TX, 83.3MHz FSB/41.7MHz PCI, it's a K6-2 not 2+ so there is no L3 cache only L1 and L2 (512kB PB on motherboard)

Video on 640x480 16bit textures

maxtherabbit wrote on 2023-11-26, 19:29:

Sound on, 430TX, 83.3MHz FSB/41.7MHz PCI, it's a K6-2 not 2+ so there is no L3 cache only L1 and L2 (512kB PB on motherboard)

Video on 640x480 16bit textures

What's your primary card ? I found an old benchmark of mine and it can vary by +30% (30.5 fps with a C&T B69000, 20.8 fps with a Matrox G450).
It's on a much more powerful system at different resolutions, but at least it shows that the primary card matters in QII.

That's interesting, my primary card is a Matrox Mystique. I naively assumed matrox+matrox would be good to go

Hmmmm, probably explains why I was completely underwhelmed by Apocalypse 3Dx performance back in 99/00 when I was trying to run it with a GD5430 PCI ... does it correlate with software Doom/Quake scores on the 2D card or is it different things which affect speed?

Here is the file
https://docs.google.com/spreadsheets/d/1p3G-l … ?usp=drive_link
Note that since then, I discovered that older drivers for the Rage 128 are both faster and more compatible than the ones used in this benchmark. Rage 128s with appropriate drivers are the best cheap cards to pair with a PCX2 IMO.
According to the error code, I had to deactivate optimizations within the drivers to run QII with the Matrox G450, which explains why it's so slow.
I recommend to favor the latest Videologic drivers that let you activate or deactivate optimizations.

A GD5430 is probably too slow. Over a certain level of performance (a quite modest one), 2D capabilities stop making any difference at all, since overclocking a 3D Rage Pro doesn't impact speed in any way. And on newer cards that's probably more a matter of drivers and hardware interoperability.
Unfortunately, I have no older Matrox card to test with the PCX2, but the G450 doesn't look good. Pairing two Matrox cards doesn't provide any benefit since Matrox isn't involved in PowerVR technology and Matrox drivers are marginally different from the reference ones.

The B69000 being the fastest on QII is interesting, it has its video memory directly integrated into its DIE, making it very fast at processing what the PCX2 sends to its buffer. Just a theory. Unfortunately, it has too much compatibility issues to make it among the best ones.
The Laguna3D and (expectedly) the ET6000 are pretty good too, but rarer and more expansive than Rage 128s.