CRAZY! Chinese Labs Double VRAM on NVIDIA Cards!

Somewhere in China tonight, a technician is heating an RTX forty-ninety's graphics chip to two hundred sixty degrees Celsius. On purpose. When it's done, the card will have forty-eight gigabytes of memory, double what NVIDIA ever sold it with. Welcome to my lab. Well, my recreation of one. Let's find out who's doing this, how they do it, and whether you can get one.

First, why bother? Because AI runs on memory. A seventy-billion-parameter model, squeezed down to four bits, is about forty-two gigabytes. A regular forty-ninety has twenty-four. NVIDIA does sell a forty-eight gigabyte card, the RTX six thousand Ada. It listed at six thousand eight hundred dollars.

In China, there's a second problem. Since November twenty twenty-three, shipping a forty-ninety there has needed a US export license, so every card already inside the country became precious. By August twenty twenty-four, the first forty-eight gigabyte versions were up for rent on a Chinese GPU cloud.

So who are they? The best documented is Brother Zhang's repair shop in Zhengzhou. Gamers Nexus filmed his team at work: orders of ten to twenty cards at a time, about two hours each. In Shenzhen, a company called CT rebuilds retail fifty-nineties as server cards on factory lines, with nitrogen reflow ovens. And the special boards? They're sold openly as kits, for about a hundred forty dollars.

So how do you double the memory? A forty-ninety talks to its memory over a three hundred eighty-four bit bus. That's twelve thirty-two bit channels, one chip each. Twelve chips, two gigabytes each, twenty-four total. And GDDR6X never came bigger than two gigabytes a chip, so you can't just swap in bigger ones.

The trick is called clamshell. Put a second chip on the back of the board, right behind the first, and let the pair share one channel, each using half the data lines. Twenty-four chips. Forty-eight gigabytes. That's the ceiling. But the stock board only has room for twelve, so they move the heart of the card onto a brand-new board.

This is where it gets delicate. The GPU gets about two hundred sixty degrees for roughly five minutes, under a metal shroud that keeps the heat off everything else, until the solder melts and it lifts away. Then the twelve memory chips come off, with the whole board sitting on a heater at about a hundred thirty degrees.

Every chip gets reballed. A memory chip has a hundred and eighty tiny solder balls underneath. The old solder is wicked away, the chips go face up into a stencil, eight at a time, and fresh half-millimeter balls are melted on with hot air near three hundred ninety degrees.

Then it all goes onto the new board: twenty-four memory chips, and the GPU lined up on top and reflowed. Zhang says about ninety-nine percent of his conversions succeed. But the failure he sees most is a short circuit, and that can crack the GPU die. There's no undo button on a cracked die.

Hardware is only half of it. NVIDIA's chips refuse firmware they don't trust, so every forty-eight gigabyte card needs a special firmware image, and how modders got it, nobody can say for sure. With it, normal GeForce drivers just work. The finished card is a loud blower, two slots thin, so eight fit in one server. And it isn't faster. You're buying capacity, not speed.

But capacity is the whole point. That seventy-billion model now fits, with room for about eight thousand tokens of conversation. And the speed limit is simple division: about one terabyte per second of bandwidth, divided by forty-two gigabytes of model, is roughly twenty-four tokens per second, at best.

So can anyone buy one? Pretty much, yes. In twenty twenty-five, Chinese shops sold them from about twenty-two thousand yuan, roughly three thousand dollars. Today on eBay, asking prices run from about forty-four hundred to sixty-one hundred dollars. And a shop in the US, GPU Lab, will convert your own forty-ninety by mail for about seventeen hundred dollars.

Before you click buy: NVIDIA's warranty is gone, and the only warranty left is the shop's. Scammers sell empty shells: cards with the chips already removed. Import duties apply, NVIDIA's driver license bans data center use, and whether owning one is legal where you live is a question for a lawyer, not an anime scientist.

And the fifty-ninety? Factories rebuild them on production lines, but still at thirty-two gigabytes. A ninety-six gigabyte version just appeared on Alibaba for thirty-eight hundred dollars, but nobody has shown one working, and the viral one hundred twenty-eight gigabyte card was a fake screenshot. Ninety-six is possible: NVIDIA's own RTX Pro six thousand uses the same chip with ninety-six gigabytes. It costs sixteen thousand dollars.

As for NVIDIA, it hasn't said a word about these mods. Chinese sources told Gamers Nexus it keeps one eye open, and one eye closed. So that's the secret: no magic, just a second layer of memory, a brand-new board, and very steady hands. Now if you'll excuse me, I have twenty-four chips to reball. Eight at a time.