Turn Your RADEON 9500 into RADEON 9700 in No Time!

RussSchultz:

I thought most chip defects were of a highly localized nature, like say, one defective gate or a short-circuited interconnect? This class of failures would seem to be fairly easy to circumvent with laser repair or bios-level configuration, disabling the part of the chip with the defect - of course provided that the chip is designed to make this possible in the first place. Are there any common defect types (other than issues that destroy entire wafers) that can't be corrected this way?

 

http://www.darkcrow.co.kr/News/Eng_News.asp?board_idx=1655

 

Or turn it to Fire GL M1:

http://www.darkcrow.co.kr/News/Eng_News.asp?board_idx=1651

 

While the result of the specific failure will be varied, failures still tend to happen at single points or regions. If you look at the block diagram ATI showed of the r300, most of the chip is covered by the graphics rendering pipes. With a little bit of design thought ahead of time, it is not difficult to make it so that the upper or lower half of the pipes are disabled. They certainly didn't design 8 different pixel pipes - rather they designed 1, and replicated it 8 times, interconnected together at certain points. When they demoed 9500Pros on the 9700 PCB, do you think it required any more than a software change to take it from 256-bit memory to 128-bit memory? Why would disabling 1/2 of the gfx engine be any more difficult?

While some of the 9500s may indeed be die with 8 fully working pixel pipes, as they ramp up production I expect most will be ones with half of the pipes malfunctioning in some way. The large area taken by the gfx circuitry makes it likely that any individual silicon defect will be in a gfx area. That puts it into one of the two halves. Since they can use software to disable one, it should be easy at test time to first check and see if the part would be a good 9700, then if not, test to see if it would be a good 9500. If not, toss it away. So they're essentially getting the 9500s from the scrap pile that would normally be thrown away - pure profit.

 

And what gives you the idea that the failure mode will leave a considerable amount of bad chips (enough to base a product line on) that have failed leaving only (up to) 4 pixel pipes damaged, but everything else functions 100%?

I don't know their design, but I do know that the number of 9500's available on store shelves has been pitifully low. (I have NEVER seen one in a store--and I've been looking). My guess is they'd rather sell 9700's, and the supply of resurrected dies (if they exist) is pitifully low.

My guess is the 9500 will die a quick death when the RV350 is born, but that's good for ATI--since their margin is likely much higher on the RV350.

 

The fact that 9500s can be turned into 9700s by reactivating the 4 pipes, pretty much means that they are not defective at all, no?

If this can be done in software it is awesome! I think the time has come for me to buy a R9500 Amateur and I never thought I'd be interested in this card!!! 8)

 

It means that the ones they played with can, without noticible problems. That's no guarantee that every one can.

The defect density quoted by fab houses is mostly a result of defects in the base (silicon) layers of the wafer. The metal layers are much easier to get right, and tend to have mistakes less often. Since individual defects (ones that do not ruin an entire wafer) tend to be small, their effect is usually on one signal or a couple of signals. While it is possible that a defect like that can ruin an entire chip, if you did the design with redundancy in mind, you can just shut off that part of the chip and you have a still functioning part. Processor vendors do this all of the time on things like cache memories - redundant bit lines or word lines or even whole redundant memories which are tested after being fabbed. The working ones are the ones selected for actual usage (sometimes by blowing fuses at wafer test, sometimes by software testing, it depends on the implementation). The theory for this is the same as testing both halves of a graphics pipe versus one half or the other, and determining which works. In the processor cache case, they build extra memories and only use the ones that work, so you might only have a 512KB cache even if 550kb of the cache might actually work. In the r300 case, if both work they enable both, if just one works they enable just one.

That depends on what the rv350 really is. If it truly is 4p/128-bit, then the only thing the 9500 would have on it would be the ability to use 256-bit memory interface. Then it would become a race between the two to see if the potentially higher core and memory clocks on a 4p/128-bit beats the lower clocked 4p/256-bit.

As for margins, for parts that really are 8p capable, turning them into 4p parts would be a loser, but for the rejects that are really only 4p capable would normally be scrap, which gives you great margins.

I guess what we really need is for someone to buy up a bunch of the regular 9500s and try out the extra graphics. If on at least one of them enabling the extra pipes results in 3d functionality that is incorrect, then they really are using failed 8p parts in the 9500.

 

No need to do that yourself.
On Rage3d a lot of folks that did the resistor modification have problems with "checkerboard" textures in 3D applications. The speed improved, but the quality now sucks. No solution has been found there yet AFAIK.

 

Well shit... I guess we'll just have to see how this goes then. If I get two gimped "9700s", then these cards are going back out the back door as fast as they came in...

 

That checkerboardedness does suggest that there could be something busted in the chip's 8 pipes.

I wonder how in the heck it knows which pipes are busted and not use those. (since apparently you can go back to the original state simply by undoing the mod)

 

Doesn't the Radeon family render in tiles (checker board like) to the frame buffer? Not sure if this is related. Maybe someone can turn off the tiling feature to see if this has an effect. I know on my old Radeon you could use a tweaker to turn this on and off.

 

Hmmm, I got my 9500 128MB card today and did the tweak, and I got the "checkerboards".

It wasn't what I expected. I expected one pixel out of every 8 or whatever to be screwed up (assuming there was a bad pixel pipe, for example), but there were 3 lines of them, and they were big (like 8x8).

Plus, the tweak only improved my score from ~9500 to ~10500.

Ah well, the card was destined for my HTPC anyways.

 

Hot Tub PC

 

Must be the noooobie soldering job <joke>

I will update my success/failure later this week...

 

Nope, just used the software version of the tweaker.

I might try again, when I pull the heatsink/fan off and replace it with something that will be silent (its a bit loud for my tastes)

HTPC = HOME THEATER PC

 

I am just going for the real thing. Hopefully a Tyan Tachyon 9700 pro. Seems like a Radeon 9700 non pro isn't that much more expensive then a 9500 Np. Each their own I take it.

 

http://www.beyond3d.com/forum/viewtopic.php?t=3859

 

Ok, I got 2 9500s today and I have to say that this hack is just BS. I don't believe any of the 9500s are working. Of the two I received, one is totally jacked, meanwhile the other has VERY SMALL errors. You have to look very closely in order to find these errors. As far as I'm concerned it's likely those with a "working" 9500 actually have one that has very small errors that they haven't noticed yet.

Also, even assuming that there are some actual working cards, my 0% success rate is a good indication that we're looking at a very poor likelihood of getting one. It's not a "30% failure rate", at best it's a 30% success rate. Either that or I have really bad luck...

Anyway, I'm just going to fork over the extra money for a working card. It's just not worth the hassle buying 500 9500s just to save $100.

 

Reading x-bitlabs turning the 9500 PRO and 9700 PRO into the FireGL cards by closing a resistor on the GPU and editing the BIOS is amazing - that is creating a proffessional graphics card costing over double the gaming card for a small change.

http://www.xbitlabs.com/news/story.html?id=1042578447

 

As suspected, Sapphire 128MB 9500 will soon be using the 9500 PRO board design:



http://www.darkcrow.co.kr/News/Eng_News.asp?board_idx=1701