Detailed information on 2 9500 mod failures

[Truckchase!]

I'm starting this thread in the hopes that I can put some concrete eveidence on the table regarding the failed 9500 mods, and to begin work to overcome the issues. If you add to this thread, please provide everyone with all the information relevant to your attempt.

First, my system specs
Epox 8k3a+ w/ 2812 bios, Athlon 2200+ @ 177x11.5 (2035Mhz), 256MB DDR CL2 all settings tweaked, Western Digital 120GB HD on VIA IDE controller, Toshiba DVD-Rom drive, Aopen CDRW, Sound Blaster Audigy Platnium, Intel Pro 10/100 Nic, 500w Power Supply, plenty of cooling

WinXP Corp SP1, VIA 4-in-1 4.45, Catalyst 3.0a, Rivatuner 12.1

Second, the cards:

I ordered two Saphire 9500 Non-Pro 128MB cards, one from newegg and one from Nicx. They both have 3.3ns infineon ram, and they're both on the exact same PCB. (Red 9700 PCB) The one from NicX *appears* to be a bit newer based off of the weathering (minor mind you... not actually "weathered") of the manufacturers' stickers on the board.

I was pretty excited to say the least, and I couldn't wait to try them out. After running an insane ammount of benchmarks on my system with my geforce 4 4600 to create a baseline, I popped the card from newegg in. I had tons of issues and eventually had to reformat my machine, but once I got it back up with XP I headed straight for 3dmark.

Before I list all the results, let me say that there is NO overclocking going on at all with these cards. I'm taking everying one step at a time.

Here's the first result with the first (from newegg) UNMODDED card.. this would just be the 9500 128MB running with the 256bit memory interface.(which is stock on these cards)

http://service.futuremark.com/compare?2k1=5638228

11264. Not bad... my high score with the GF4 (without the lod CHEAT) was 13635. After that I proceeded with the software mod. Using the catalyst 3.0a drivers, I ran the patch script and re-installed the drivers. Upon rebooting, I ran 3dmark again.

Checkerboards. Damnit. Checkerboards right across the middle of the screen, and only there. They seemed to be spaced by about 32 pixes, and the boards themselves were 16x16 pixels wide. (cubes) They were present in EVERY test, but the card was totally stable other than that. Here's the 3dmark result:

http://service.futuremark.com/compare?2k1=5638418

Awsome score, 14035 before I've even tweaked my system or overclocked the card. Obviously tho, I can't use this with checkerboards on the screen at all times. Now, here's something that I want to point out. Look at the multitexture fill rate of the card.. unmodded vs. modded. This is how to tell if the mod worked successfully or not for sure. I'll explain this in further detail below, but for now, look at the fact that unmodded the fill rate was 1104 vs. 2173 for the modded. Almost double, as it should be with double the ammount of pipes and memory bandwidth to spare.

I tried both turing off "Write Combining" and Fast Writes and a combination of either/or to no avail. The checkerboard pattern wasn't affected at all by either. I also tried messing with the AGP arperature. That also had NO effect.

Frustrated, I popped in card #2. This is where it gets odd. I had the hacked drivers already installed, so I just kept them that way. Upon running the benchmark I was excited, because I got no artifacts at all. It completed the benchmark successfully, but when it finished the score turned out to be right between the old modded card and the old un-modded card. Perplexed, I put on stock drivers and ran the benchmark again. Here are the results of each.

UN-Moded: http://service.futuremark.com/compare?2k1=5641781

11497

Modded: http://service.futuremark.com/compare?2k1=5638593

12765

compare that again to the first card modded:

http://service.futuremark.com/compare?2k1=5638418

Weird eh? I tried several different ways to mod including Wizzards' mod, and they all turned out with the exact same results. So, unmodded, the two cards performed the same, and without error. With the mod, the first card ran at full speed but displayed the checkerboard patterns, and the second card ran without artifacts, but at a crippled speed.

Here's where I want to touch on the multitexturing fill rate again. Notice on this second card, after the mod, the single texturing fill rate went up about the same as the "successful"(w/checkerboard) card, but the multitexturing fill rate didn't go up AT ALL. On the other card, the multitexturing fill rate DOUBLED. I'd like to focus on this figure a bit, so if anyone posts follup results on either a success or failure,[/b] please post your single and multitexture fill rates.[/b] Thanks

Let me re-iterate that to the naked eye, (and as identified by several hardware utilities) these cards are EXACTLY the same.

This is weird. I've been unable to come up with a concrete theory( ) as to why these cards would illustrate this behavior. I thought perhaps I had screwed up a driver setting or something (even tho i didn't touch the control pannel) and deleted all the driver settings from the registry and re-installed. Same result for both cards. I chaged the order of testing. Same result. I tried different driver revisions. Same result. I re-formatted and re-installed the OS. Same result. I tired DX8 vs. DX9, screwing with all the AGP settings, different motherboard bios revisions, and doing a rain dance. All same result.

Now for the questions:

Let's theorize as to what could be causing some of these chips to fail in either way.

Let's theorize as to what could be done to complete the mod successfully on these cards.

I'm hoping if we get a discussion going based of off concrete info that we can find some way to address the problem. If you have had a similar experience, please share with all relative information. If we can figure out a trend on successes vs. failures, perhaps we can figure out a way around these issues.

I'll be trying a volt mod on the cards tonight, and I'll post what happens here as soon as I'm done testing.

[funnyperson1]

it sounds to me like the second one is based off of a 9500 pcb, and in effect is a 9500Pro with a 128bit memory bus, that is the only way i would think you would get these results

[ByteWipe]

my brother just did the mod on his 9500np 128mb model..... he didn't get any artifacts, got about a score of around 12500 (XP 2100+, not o/ced)

Still though, it doubled his performance from the old specification....

[PackardBelltoll]

I would get the bios from the card that had artifacts but benched well and put on the card that did not have artifacts . Make sure to backup bioses you wish to keep. That bios from the card could make it better. I know a bios can change driver performance.