New meaning of "eGPU"

[Khenglish]

So one thing I find fun is to make older hardware capable of running modern games at decent fps. The most recent target was the laptop I started college with from 2008. It had an mxm 2 slot build for mid-size mxm cards. mxm 2 was dropped for mxm 3 a few months after I got the laptop, making upgrade options very limited. Considering that the laptop was virtually worthless due to the age, to widen my options I cut out the left side of the laptop and removed all motherboard connectors (VGA, HDMI, ethernet) and associated circuits in the area to fit in a full size 256-bit mem card. I got a supposedly inoperable mobility 4850 for $31 off ebay with the idea that I would fix it. I popped it in, drilled some holes in the gpu heatsink to match up with the wider ones for the bigger card, and got this below:

http://imgur.com/FeUqkmS,XkfecNw,CQkmegR,XKtI92c,SPnBFS6#0

So yeah, it worked. Yes, the card sticks out the side of the laptop a few mm. The seller was wrong and sold me a fully operational card. The laptop BIOS properly recognized and ran the card, and did not flip out over the 2.5x higher tdp. On the screen in that picture is the final graphics test in 3dm06. With a giant fan blowing in under the laptop the temps were good. Good enough in fact to allow overclocking.

The mobility 4850 is just like the desktop 4850, just lower clocked. 800 stream processors at 500MHz with 1.05V, and 512MB 256-bit GDDR3 at 850MHz with 1.8V. The desktop 4850 runs at 625MHz at the same voltage, so I figured I'd max out overdrive at 550MHz and see how it scored, then figure out the stupid 550MHz lock later. After about 10 seconds into the benchmark it crashed.

I was very confused. Crash at 550MHz? I googled the desktop 4850 and saw people were getting mid-high 600s, one even 700MHz stable at the same 1.05V. There's a mobility 4870 on hwbot at 680MHz. After further testing I determined my max overclock to be 525MHz on the core. Something had to be wrong. My core could not possibly be that atrocious. I then proceeded to upgrade the voltage filtering to match that of desktop cards:

http://imgur.com/FeUqkmS,XkfecNw,CQk...I92c,SPnBFS6#2

Above is the inductor swap. The pcb was built for larger inductors on the core, and had room for a larger memory inductor as well, so I upgraded that too because why not. The core went from .36uH inductors to .56uH. This lead to a 36% reduction in current ripple and thus voltage ripple to the core. The result was a 15 MHz improvement in stable core clocks. I replaced the .68uH memory inductor with a 1uH.

While 15 MHz was nice, I was still 100 MHz short of having a decent core. I then improved the filtering more:

http://imgur.com/FeUqkmS,XkfecNw,CQk...I92c,SPnBFS6#3

The core had 2 1000uF electroyltic caps to filter out voltage ripple from the buck converter. While this is decent, there was a lack of ceramic caps for removing high frequency ripple. I could only find 4 ceramic caps for this purpose, meanwhile I noticed my other newer cards had 12 for this purpose. You can see where I added 8 ceramic caps near the core inductors. I also increased the electroyltic capacitance by 3x 470uF caps. One is sandwiched between the 2 1000uF caps, and the other 2 are stacked on top. For good measure I added a 470uF cap to the memory's buck converter too. This resulted in a 40 - 45% reduction in ripple from the buck converter, and an unknown reduction in high frequency ripple from other sources. Combined with the inductor swap, ripple was reduced by 60-65%. The max core overclock improved by 20MHz to 560MHz, for a total improvement of 35MHz, or 7%. There's likely another 5% left, but the filtering is now on par or close to well designed desktop cards, and the core clock is still garbage.

I noticed that the core was too big for the laptop's original gpu heatsink, not touching the outermost 1mm of the core around 2 of the edges. I figured this was insignificant and would only help memory clocks in anything, but I figured I'd replace the heatsink contact with a copper plate from a Clevo heatsink. A little toaster oven time, hack saw, and solder paste later the job was done. MEMIO temps improved a little, but no overclock improvement as expected. I inspected the paste and verified that I had 100% die contact. The "new" heatsink is below. (There's AS5 all over the top of the CPU heatsink because I used to open the back panel and slap a Pentium 1 heatsink on with a big fan to run the 45nm C2D at 3.8GHz 5 years ago.)

http://imgur.com/FeUqkmS,XkfecNw,CQk...I92c,SPnBFS6#4

So yeah, there are RV770 cores 25% better than mine. Max temps with the overclock is 72C on the shader core, which is not high at all, and the core clock sucks. Filtering is now in good shape, and the core clock still sucks. I went through all that to get the card running and try to fix clocks... only to get a complete dud. It looks like the mobility cards got all the terrible cores that could not meet the desktop 4870 or even 4850 specs.

So the plan now, not necessarily in this order:

1. Replace all the internal radiators and add a heatpipe to see if that's enough to get the internal cooling system good enough to not need an external fan. It likely won't be, so I'll try to find a slim 2nd fan to put over the gpu. I'm not sure how I'll get a radiator in there for the 2nd fan yet though. Alternatively and more simply I could get a deeper single fan, because this isn't going back together properly anyway. Big laptop fans are expensive though, and desktop GPU blower fans seem too deep.

2. Replace the garbage core. I think I may be able to get 1-2 dead mobility 4870 cards for free. Fix them and test in my main system (the cards and laptop are mxm 3). Figure out which has the best core. Remove that core and reball. If I do that without killing the core (the packaging bubbles up and bleeds if you do. easy to spot), assume I still won't kill it by putting it on the mobility 4850. I'd put a desktop core on, but when I did that once for an ancient mobility 9600, only the non-LVDS display outputs worked (not routed on desktop core due to LVDS not being used). Since the only display output left on the laptop is LVDS due to making room to fit the card, I am hesitant.

3. Figure out some way to close the laptop back up while being structurally sound. Lots of ?? when I try to think of how to do this. Temps first without giant external fan since that affects this.

4. GPU power draw increased by 2.5x. When the mobo's power-in FETs blow, find and replace them with something stronger. Maybe find them beforehand and double stack the current FETs.

5. Try to find modern GDDR3. Drop card voltage to 1.5V and swap the chips in. This is lowest on my priority list because of the time reballing 8 chips would take, modding the BIOS to recognize the additional memory, and questionable performance improvement as the existing memory already does 1100MHz, which is about as good as I have heard of for any GDDR3.

So has anyone ever gotten such a relatively terrible clocking core before for anything? 25% max clock variation between identical parts is something I have not heard of since the 90s. The terrible core clocks are an incredible let down on such a crazy and otherwise successful mod.

Max stable 3dm06:

Generic VGA video card benchmark result - Intel Core 2 Duo Processor E8435,COMPAL JHL90

I had 10600's with an unstable core overclock. CPU is not overclocked because doing so would cause the system to shut down with the power hungry gpu. If the core was decent I would be over 12000...

Vantage gpu score is 6600. Again high 7000s, possibly 8k with a good core...

[Ralph]

Hi!

Great work. I have the same "problem" with my Ati 4850m that I have in my msi gt725. I can overclock to stable 530mhz and if i try higher i get random freeze for few seconds not bsod but it will freeze and after that will work ok. I think the reason for this is because we have low gpu voltage only 1.05v. Desktop cards have like 1.10v.

What I did was extract my vbios using gpu-z and change gpu voltage with RBE. I tried bumping voltage to 1.06v or 1.07v and reflash my modded vbios. It flashed with success but I had to reinstall my ati drivers for changes to work. Unfortunately after that when I tried to use any 3d game i got random artefacts on the screen ( black random lines ). So I went back to my stock vbios.

Maybe try rising gpu voltage in your vbios? In my case this did not work but maybe it will work with yours gpu?