I Doubled My GPU’s Memory (in a guy's basement)（原声文稿）

Transcript of 把一张显卡的显存强行翻倍：地下室硬核改装实录 （中文完整文稿）。

Two flights and eight hours of work later, I have accomplished it.

I have two 48 gig NVIDIA GeForce RTX 4090s, but why?

The monitor on your desk is increasingly just a window into somebody else's computer.

Your files live on their servers, your software needs their permission to run, and your AI agent, which is of course, trained on everyone else's work, takes either a monthly fee or you get gouge for tokens, and then on top of that, all of your data is stolen and either sold to advertisers or used to serve ads directly through the app itself.

In short, the man wants you to forget that the P in PC stands for personal and a lot of people are getting fed up about it.

There's a massive surge of interest in topics like Linux, self -hosting and the focus for today's video, local AI.

For the past few months, I've been scrounging up all the hardware I can get to build a sweet little, we have clawed at home machine for myself and my family but the same companies building AI for the cloud are buying up all of the hardware, driving prices through the roof for the rest of us. So I'm getting a little scrappy.

I got insane deals on a pair of 4090s and I'm going all the way to Michigan to upgrade them.

Not replace them, upgrade them from 24 gigs to 48 gigs each, giving me 96 gigs of VRAM across both cars, allowing me to run some pretty massive models that would otherwise take professional cars that would retail at like 15 grand to be able to run.

To be honest, I'm a little scared, But Greg the man behind GPU lab will be doing all of the hard work. So Greg where do we start? So Luke you sent in your cards ahead of time.

Yeah, and basically I had them drying overnight and And we're going to take them out of the oven right now.

And basically, here's your, that's hot actually. That's really hot. No, I got used to it. I should have told you, I should have told you. No, that's fine, that's fine, I'm good.

So why do they need to just try them out for moisture and stuff like that just to prepare them for the process?

So long story short, we're taking boards from China, adding some memory modules from China, bottom bulk on a tray.

We're taking the memory off of my boards, taking the core off of my board. These are two different cards.

There's like a founders edition and I think it was a gigabyte board and then transplanting all of that onto that That Chinese PCB.

Yes, correct So that's the like way too simplified version of what we're gonna be doing for the next few hours. Yes I noticed there's like a million cards.

Yeah in the oven as a small shop I'm doing typically maybe a dozen or two cards per month since my exposure with Wendell Things have kind of kicked off and I've had to purchase a lot more cards So you're not just upgrading other people's cards.

You're also selling pre -upgraded cards. Yes. As I can get them, they're a little bit difficult to get. And if you go on eBay and buy a dozen of them, that will drive the price up. So you have to like kind of trickle by them.

And I would honestly prefer people send me their cards because that outsources that hard part of me trying to source cards, test them, just focusing on upgrade is a lot more easier for me.

Essentially what the design is, is a copycat of the 3090.

Basically they sandwiched two memory modules on top of each other on each side of the circuit board on the bottom and top side and those memory modules on each channel share the same data pins and the address pins and the GPU core basically does like a chip select like hey I want to address the back side or the front side really quickly and so to the core it could basically address 48 gigs of memory yeah not more than 24 at a time but it could do it so fast that it doesn't matter so I guess next step we could take these and start applying memory modules to the rear.

We want to basically apply thin layer of flux and then put the memory modules one by one under a microscope. I look under a microscope just to make sure all the solder bumps are flattened.

The pads come from the factory pre -tinned with lead -free solder. It's like a protective shell.

And so one of the process steps is to actually go on with a solder wick and an iron and basically flatten all of the pads.

Once you get them flat, then you can put the memory modules on.

But sometimes I miss some pads and they're just, you know, half -flat or a little bit like have a little point on them right and that makes putting a memory module on there a little bit difficult because it goes off skew just getting them completely flat sometimes is a two -step process now we could put the modules on okay so when you're lining the chip up I'm noticing that you're pausing so I'm doing a visual inspection of the BGA spheres basically just making sure there's no foreign materials there making sure all pads have spheres and there's no connections that shouldn't be you've placed all the memory modules on the backside what do you do now we got to apply them to the circuit board now.

Oh, we gotta do one side at a time. So we're gonna start with the backside.

This is our BGA rework station. This is used to apply memory modules to our new circuit boards. Technically in China, they do it one by one. You could do it on the same machine that we would extract memory models from.

But here I have special nozzles that basically allow it so I could solder on 12 new memory modules all at the same time under the same reflow curve. So it makes the process a lot more consistent and repeatable. You're also gonna use this later to remove the cores. Yes, correct. I get a side view with this microscope camera and a lens.

The lens, I don't really know what it's for. I think it's like from a scientific application. You don't get like a perfect image most of the time, but this is good enough to see the balls melt, like when that time is.

There's no way when that time is crucial.

It works good enough, but the problem with high magnification to do something like this is you start getting like a lot of aberrations when you zoom in a lot and then also needing a lot more lighting.

So it looks like the process is finished.

So now we're just gonna like lift the heat, stop processing and start the cooling.

So these PCBs from from China you were mentioning that they had Micron chips on them before But Micron isn't making those anymore.

No about December They stopped the whole market dried up and it was very chaotic all the third -party sellers of new old stock completely They clutched all of their memory modules and they would listen for ridiculous prices Then people started buying them for those ridiculous prices So you could get the new old stock for a while and then you couldn't and so now we have to use Tested memory modules from boards that have bad cores, but they pull off the memory modules They validate them through a memory testing fixture like this sometimes to get the boards with the memory modules and sometimes without and for yours I did it without so we just kind of see the whole process.

So here's your board So now we're gonna do the exact same process we did for the first one Okay, I probably have you do the flux all right.

Yeah, and also start putting on memory modules I probably do like half of them just because even my little brother I started having to him He was like super slow at it I was I was slow too when I started it might take you a minute to get used to like the microscope and just like having that go. I can't see anything. So adjust these like this.

Wow, that's not the easiest. Great, so if you want to take some flux. Like that?

Yeah, just kind of like sparsely like I did, not like, yeah, just like the spotty. Yep, that's perfect.

Okay, cool, set the side and then you got your paintbrush. Is there a specific way to stroke or anything? Just as flat as you can.

Don't try to have it like pull up on like the edges or the corners anywhere. That lone rogue one on the side. Okay, great.

So now you have the pads flexed and now we're going to take the memory modules that I've pre -prepared.

One by one, we want to start putting them on the circuit board.

This has been quite nerve -wracking and I think I'm really slow, but it also started to feel a lot better as I went along.

This is the last one that I'm placing now and then we'll get to evaluate how poorly I did. Greg. Yes. How'd I do? Excellent.

Really? Yeah. You didn't even look. I know.

There's one misaligned, two misaligned, but honestly like that's within spec. I'll take two.

Let me just poke at it. Like it honestly would flow into place.

I'm just OCD about it. I don't want to redo this. It can be like a 40 -minute detour. Yeah, wow. You got progressively better.

Yeah. These are like your first ones.

Those are where I started, yeah. And then like the other ones are like basically how I do them.

Nice now we just got to start the process turn off cooling Luke go ahead press start Awesome look mom. I did it. So probably around now.

We might start. Oh, yeah melt.

I just saw this one go Yeah, so it's like around 240 that they start going but when we actually melt the GPU cores themselves You'll see them dramatically like flow into place and it would be a lot more cinematic.

I guess Now it's kind of doing it's cooling down thing pretty hot right now so it feels like I'm still seeing things happening but that's just the flux I guess those balls are still melted it's still pretty hot but basically when things cool down you'll see the lead actually recrystallized sometimes yep so this one's ready now we're gonna take off the core and basically I had to change out the nozzles these ones were for applying membrane modules so this is the gigabyte card the gigabyte card was given to me by mr.

dank pods Wade will we'll see if it survives what we're really waiting for is the GPU to like Pull off the board is that correct?

Uh, yes We're about the temperature right now where the solder balls are almost liquid us so you can see like the solder balls like starting to stretch I'm like lifting up and down manually So it's like almost ready to pull up.

It looks like there's like surface tension now holding it It looks like and I'm gonna try to break some surface tension right now Yep, there we go. Okay, so now we pulled off the chip and stop it start cooling.

Oh my god the chips ready now we got to take the circuit board and then put it on the station to start taking the memory modules off now I want to put a thermal couple on the board and that's how the IR preheater kind of knows how much power to radiate into this board this is gonna start putting out a ton of like like a kilowatt plus of IR energy and just heating up the board the boards are made black specifically for the reason of helping reflow basically during the process of when they actually make these boards.

They use IR heatings I guess like a conveyor belt that takes them through stages of infrared heating and that kind of mimics a temperature curve profile and so a black solder mask is a lot better at absorbing infrared heat and it heats the board more evenly.

I'm not gonna lie for years I thought it was just because it looked cool.

Now you're applying heat to the individual memory modules in order to lift them. Yep so we're just going and pull them off one by one here. So that board is clean. Correct. No GPU, no memory modules.

And if you were a lesser person, you would take this, put it back together with a cooler and sell it on eBay for like full price.

I have seen a bunch of that.

Yes, Facebook, Mark and Place do not buy GPUs without any kind of buyer protection. So now we have the second 4090 that we're gonna upgrade.

We basically did the process of taking off the main GPU core from your Gigabyte one and all of the memory chips.

These are gonna be set aside. We're gonna take off the chip and the memory module off this one as well now.

NVIDIA supposedly very cheaply makes these, so they're like susceptible to over -rework. It's not as shorter than me. Oh yeah, yeah. Seriously? Oh yeah, yeah.

He looks like a giant on camera. So here we have 12 memory modules from one of the, if it was the gigabyte donor card.

We basically need to remove the solder bumps, and we need to make them flat.

The first step would be to apply a bunch of flux to the top of them, and this would just get it ready to make it easier to wick up all the solder from it. Just goopy with it.

I'm just gonna go over each of the pads.

And that's just removing excess material Yeah, if I went straight to the wick the wick would get so saturated.

I just waste a bunch of it Yeah, so just going over Let's me get each pad like definitely clean if it didn't have any solder on it I like it it pre -tins it basically make sure it's electrical or physically. It's gonna bond There's like three stages. There's the raw chip with a bunch of solder everywhere.

That's gonna be really messy And then there's this step where they're like little bumps probably like like that like a little bit of a sphere And then we want them totally flat.

That's the final stage We're gonna put flux and then spears on there and then the spheres are gonna adhere to it And we're gonna have basically fresh chips, and you're not worried about like scraping them or anything Not with the lint -free cloth Well, I'm seeing like every once in a while a chip will overlap another chip Yes Don't seem to be a major concern as long as you're gentle and you don't force down when the chips like on top of another Okay, so I remember the first step was just like copious flux. Correct. There was tons.

Do like three lines per row of chip It looks like he was putting caramel on a dessert.

Okay, you want to take the solder line over here?

You'll see you're wicking up excess solder Yeah, and some of the pads that look like they're flat already You'll want to kind of go over those a little bit just to make sure that they're tinned They'll be easy to adhere to.

Oh, yeah, okay Don't stay on one chip for more than, I'd say, 20 seconds.

So it looks like Luke's done with cleaning up the pads. Now we're going to switch to Wick.

Trying to get the correct soldering iron position, trying to make sure that it moves along, and keeping everything in view in the microscope, is kind of a nightmare. Oh, that one went great. What the heck, what changed?

I don't know what I did differently, but that last pass was like, it felt completely night and day to every single other one I've done. There we go. I'm worried about some of the first ones that I did. Great, you can start next Monday. Let's go.

I'm just spraying some distilled water over them because the flux is water soluble.

Now putting on IPA or isopropyl alcohol just on top of the distilled water really.

So they have past inspection and we are framing them so that they can be ready to to receive their solder balls.

So we have Luke's two cores and now we're gonna basically flux them up and start cleaning the solder bumps off of them and preparing them for a re -balling process.

What is flux? Basically, when you cook, you want to cook with cooking oil.

It kind of serves the same purpose. It allows heat to flow.

I think it first started back in the day with like pine rosin. Oh. And then chemically, like just like oil got synthetic.

This stuff got synthetic too. I think some of it is still rosin based, but this stuff isn't.

What it actually does is that it creates a barrier for oxygen not to get to your solder So it doesn't oxidize it while you're working on it It helps heat transfer from the iron to if it's wet and there's like a lot of heat It flows through the flux really well rather than the air air is a really good insulator.

That looks insane So we have the chips in the fixture now And we want to basically spread a bunch of solder spheres all over the pads just on the pads and to do that We have this foil here foil stencil.

They're called multiple things we put this over top just like this we want to line up just like that and it fits into grooves.

So this won't move right now and this will allow me to take correct size solder spheres and basically this is where like macro would go in.

I'm looking for holes without solder rolls.

Sometimes they pop out of the stencil and it's just easier to place them manually under microscope.

Just do like three four visual inspections just to make sure there's no extra spheres missing spheres or anything connecting so this looks good to me it's good to go into the oven and that oven you're you're cooking enough to attach the spheres but not like flatten them out yes correct it's an infrared oven so there's no air blowing or anything and so the chip absorbs that infrared energy from those heating elements and the heating elements follow a pre -programmed temperature curve that goes up lets the heat soak and then keeps going until that reflow temperature that's peak reflow temperature for like 20 seconds ish or something and then it cools it back down.

You want to do it proper and do like a slow cool down so you have a certain crystalline structure.

Sure. We have like one that like merged with another ball.

So now I just sucked up that double ball and I'm going to clean it up and I'm going to put another few sets of spheres on it because it's still hot.

I can still rework with it and it'll be fine. We have our second GPU core in the oven right now.

It's being reflowed And we're going to start with the memory chips. We have these all laid out on our holders.

And we're going to dab some flux on them and then put the stencil on and then...

So same idea as what we just did with the cores but now with all of the chips.

Is it different size ones for the memory in the core? Yes. It is like 0 .05mm difference. It's 10 % smaller.

We have the memory modules that are just finished doing their reflow profile.

And I noticed a few spots that had some solder spheres slump, so we're gonna correct those real quick But in general these are gonna cool down for the next few minutes I'm just gonna work on the ones that kind of stuck together This is the only time the audience sees this kind of process like I hate to imply that there's always issues But it's not like always like that. They're going into the fryer.

Yes a mini fry Yeah, so we have to clean off the flux just off the chips before we reflux it on the board So it's not actually frying them.

It's an ultrasonic cleaner It's still they don't get lost they're in little individual baskets This goes in here for a few minutes. It's not water, right?

It's water distilled water So now what they just finished cleaning and I'm gonna take them out.

So get your hand out receive them and Then go put go put them in the oven so Luke's gonna now flux the pads for the memory modules and Yeah, get them ready. I think I'm gonna get a job. You're doing great Luke.

Yeah Yeah, it is extremely cool to see what people are doing, to breathe new life into not even that old cards, doubling memory, all the stuff going on with the HX170s, the BC250s, there's a lot of hardware hacking, using things in ways they weren't originally intended to be used, upgrading things, etc.

It's unfortunate that a lot of it is so high priced that people can't afford it, But it is cool to see people doing this with hardware.

It's very fun, but very expensive time to be into computer hardware. Let's get out the memory modules. Do you just like pick them up? Yeah, just in your hand.

And you can tell they're really hot, but they don't have a large thermal mass.

Yeah. So it's okay to, I guess, leave it for about a minute rather than longer, like I would a core, and then they're ready to apply. See, I am the GPU now. He's even using tweezers to play some memory.

I am back at placing the memory modules so I'm looking for this little dot in the corner and I'm lining it up with something that's really hard to see on the board.

There's a corner of it that's, I don't know, open, broken, whatever you want to call it. So then this chip needs to go like that. Boom. That was actually a good upthrift.

Now I just have to do that ten more times. How many of these do you think you've done? You sell like 200 and something? Yeah 200 since the start of the year.

My thoughts are this is that's I think probably the easiest part of this whole thing and it's hard. You want to grade me sir? Okay sure thing. Let me know. I'm sure you did great.

Great, great, great. Yes like on the corner here there's just tiny like little bumps.

Yeah that one really felt horrible and then yeah those four just bump bump bump bump done. So do you want to pick up this board and we could take it to the BGA station? Yeah let's do it.

Put yep like that and then let me move that rail into place perfect in place good so now I'm going to lower this thing so it's a little bit off -center to the right right now just because we had it for the core extraction instead and we're starting to feel pretty real yeah yeah oh man our solder balls just melted our temperature profile basically hit our peak reflow and now we're done with this flow run and we just want to hit stop and go manual cooling lift the heater off of the chips and just let it sit there for a minutes.

The second board is going on the BGA machine, the first board, 24 gigs of memory this side, 24 gigs memory that side, and we are moving on to the core. Yes, correct. So this is the core. Okay.

And we're gonna put it like this. We're just going to put on just like that. And it goes on the BGA machine.

The BGA machine comes down, picks it up, and it optically aligns it with the special prism that pops out.

Okay. And once it's aligned, then we flux it up, put the chip back on, and run our profile.

So So this is cooled off, it's cool to the touch. I'm just gonna lift off the nozzle.

Give me some room. Luke, you wanna take this? Yep. Great.

Take off the nozzles that we used to apply memory and put on the nozzles to apply the core.

The vacuum pickup nozzle is gonna help us pick up the core before we apply it to align it properly. I have a laser alignment.

Basically, just like this.

This gets me relatively to the center.

turn on the vacuum pump and realign it visually. Just picked it up.

Gonna switch our HDMI input to the prism camera.

So this screen shows that the spheres which are the smaller white circles are a little bit off to the right. Yes Luke, you can notice it. Great.

So to adjust that we have these kind of really small micrometer twisty turning knobs.

So if you want to like kind of just this center area of spheres try to get those aligned on top of the pads.

Okay, so we've done this center part, but I also go off to the edge kind of like that and you can see these ones are also off Tilt so this is called a theta adjustment and this is adjusted with this one right here.

Okay, I mean just go back double -check this Doing that data just hadn't really shifted those off. So we're good to go now before we lower this on We're gonna apply some flux Luke. Do you want to lower it down now?

Yeah So pull it down. It goes automatically. It's gonna stop about a few centimeters up So, stop the vacuum pump.

So those are in focus and that was released, so I'm just going to bump it up like a few millimeters and then I'm lifting the bottom nozzle up just to meet the circuit board. So we have like an even flow of heat. Now we want to start the profile. Do you know how to do that? Yeah.

Yeah. Awesome. I did it.

So now that that started 5 minutes and 40 seconds left and we'd have to throw a cooler on it and test it, but like the card's effectively done. There it is, Luke. Oh, crazy, man.

Like it looks like we just pulled a cooler off of this and it's like, hooray.

Okay, well, I guess we gotta put a cooler on it now. Yeah, we gotta do the reverse of that. Luke, you take this board and I want you to do the most satisfying part.

Snap off the piece for me. Oh my goodness. Okay.

Oh, that does make my soul die a little bit.

So these are the coolers for the 48 gig 4090 they look all blue and everything will look like Intel but no they just have basically covers on the thermal pads that come and yeah just have to take them off.

It is way smaller than the cooler that it comes with yes and these are loud.

Yes they're made for density and because of that and the 4090 still pull 450 watts in this configuration so you need a delta fan that moves at like 5 ,000 rpm. So it gets loud.

Yeah, so you also offer water blocks.

People might want to water cool so that it's not so loud.

But if you're putting it in like a data center type environment then who cares.

Now I'm gonna take off the pre -applied thermal pad protection and we're gonna test the card like right away. Okay. The board's still a little bit warm. It is still warm.

The heat sink is going to take care of that. The residual warmth. I would hope so. It's going to be fine. Let's take the board.

Got to plug in the fan connector. Makes sense.

OK, great. Cards ready.

I feel that it's cooled off, so I'm happy to plug it in and turn it on.

So this is the display port cable that goes to this monitor right here. This is our waiting for VGA LED, the orange one. Once that goes away, we're going to get an image. Oh, let's go. It's posting now around Windows.

So this is the program called Vulkan MemTest.

It's available on GitHub for free. Anyone can use it. And it's one of the tests I do to test my cards.

We just let this card go for about five minutes doing this Vulkan MemTest. See the temperatures. See if there's any errors. One little note. We see 850 gigabytes per second.

The max theoretical rated for these GDDR6X modules is about 1 ,005 gigabytes per second. You can see more on Linux.

On Windows, it's like maybe 30 to 50 lower.

But when it goes under a full stress load, You see around this like 800 like 900 to 850.

So it's perfectly within spec It's identical to an unmodded 4090.

We remove the cooler remove the thermal paste and now you got your card And you want to basically dunk it in the water Which is a counterintuitive you would imagine Yep, that feels horrible. Okay.

I put in fresh water.

Yeah, you could see the flux already kind of clouding out of the edge of the They kind of like watch it come out of everywhere. Whoa Oh! Luke, I want you to help me with this part.

Yeah. Essentially, you want to be blast in water under the BGA to get all that flux out from over there. So hold on to the hose.

Yeah, so that went for about 11 minutes. Hold the card out and I want to air dry it real quick. Yeah.

Great, awesome. So now we just got to put this in the oven for about 12 hours and it's gonna be good to go.

Second card passed, they are both going in the oven. The first one's already in here.

They have to sit in the oven for 12 hours.

so I'm not actually going to be taking my cards home. Greg and you'll be shipping them to me later. But that's it, we're done. They both work. You did 99 .9 % of it, but how'd I do?

You did awesome, Luke.

I'm glad to have you, I'm glad you came by. You're honestly a great set of helping hands. So you picked up the work right away. Thank you. And honestly, Luke, welcome to the team.

Dude. Excited to see you next Monday. Honorary employee. Yep, start work. I'll take it.

Thank you so much Luke. No, this was sick Thank you very much.

That was that was fantastic Just so you guys know you wouldn't be coming out here and doing this you would just ship them your cards And you offer a bunch of different services.

Yeah, so we started upgrading 4080 supers from 16 to 32 gigs Do a 100 conversions but those are less known and then also 170 HX mods are coming because there's so much in the works with that modification right now, I'm not offering it, especially with the 48 gig stuff, I'm focusing on that first. The One 78Xs will come just a little bit later. Thank you guys for watching.

If you wanna see more of this, maybe check out the 170HX video that Linus did and I'll see you next time.