On September 2 a customer who runs an interior design studio brought a desktop into the store. The card was a GALAXY RTX 2070 he bought in 2019. He had been playing Black Myth: Wukong heavily, and the screen had started breaking into snowflake-like patches of color before the game crashed to the desktop. His 3DMark score had fallen from 9,800 to 5,100. He had read enough online to believe the diagnosis was a cold solder joint on the memory and that "it is cheaper to replace than repair", so he came in to ask whether the card was worth saving.
Step 1: Identify which kind of artifact you have
Screen artifacts are not one fault. There are at least three, and the repair path for each is completely different:
| What you see | Most likely cause | Can it be repaired? |
|---|---|---|
| Regular color blocks or a checkerboard pattern, appearing at a particular viewing angle, with benchmark scores falling steadily | A cold joint on a memory chip, or one damaged chip | Yes, by reflowing the joints or replacing the chip |
| Full-screen snow, artifacts that worsen as the card warms, sometimes boots and sometimes does not | Fatigued BGA solder joints on the memory | Yes. This is the case here |
| Driver error Code 43, a darkened area over the core, no display at power-on with the fans at full speed | A damaged GPU core or a blown power stage | Rarely worth it |
We start with the non-destructive checks. Pull the card, clean the gold contacts with an eraser, seat it alone in the first motherboard slot, and clean-install driver 566.36. Nothing changed, which ruled out contact and driver problems. Then we ran memory testing along the lines of MATS: on NVIDIA cards, an OCCT VRAM test plus a 3DMark stress loop. In the third round, with the core at 71°C, snowflakes appeared, and they disappeared once the card cooled. Temperature-dependent, reproducible, concentrated in one part of the image: that is the signature of a fatigued memory joint.
Step 2: Reflow the BGA. The temperature curve is the whole job
The memory chips are BGA packages, with the solder balls underneath. The fault is a micro-crack between a ball and its PCB pad. Heat expands the package, the crack opens, and contact fails. Reflowing means melting the solder again under control so the crack heals. Two things ruin this job: not enough heat, which achieves nothing, and uncontrolled heat, which kills the core or blisters the board.
This is our sequence. Do not try it yourself without an infrared thermometer, because you are otherwise heating blind:
- Strip the cooler and clean the old paste to expose the PCB. Tape off the area around the memory with high-temperature tape and wrap the core in foil to reflect heat away from it.
- Let flux wick in. Put a little BGA flux (AMTECH NC-559) around the suspect chips and let capillary action draw it into the ball gaps. Do not flood the board with flux remover (board cleaner)., or you will push the flux back out.
- Preheat. Hot air gun on setting 3 at 180°C, 15 cm from the board, 90 seconds over the whole board, until an infrared reading confirms the back of the PCB is at 80-90°C. Skip this and the sudden heat will warp the board.
- Reflow. Switch to a small nozzle, raise the temperature to 320°C, hold 2 cm above the memory and circle for 25-30 seconds, until the surface reads 235-245°C, the reflow window for GDDR6 lead-free solder. Then let it cool naturally for three minutes without touching the board.
- Do all eight chips, six on the front and two on the back. Betting on which one is bad is a worse strategy than remelting every one, and the extra work is 20 minutes.
- Clean up. Anhydrous alcohol and an anti-static brush to remove flux residue, then dry.
Step 3: Reassemble and prove it
The core gets Shin-Etsu MX-7921 paste, which had dried to powder, and the memory gets 1.5mm pads because the originals had collapsed. Then we accept the job against four tests:
- 3DMark Time Spy back above 9,700, from 5,100 before the repair.
- One hour of OCCT VRAM testing with zero errors.
- 30 minutes of FurMark with the core steady at 74°C and no artifacts.
- Two hours of Black Myth at 1440p on high settings with no crashes and no snow.
The machine then ran a four-hour continuous burn-in before delivery. Total cost: RMB 200 in labor plus RMB 80 in pads, paste and flux, so RMB 280. The customer did the comparison on the spot. A new RTX 4060 with similar performance is RMB 2,300, and a used 2070 still runs RMB 900 with unknown history. The RMB 280 was easy to justify.
Cards we will not repair
- A dead core. Code 43 plus a short on the core power rail means a success rate below 20%, and even a success tends to come back. We advise against it.
- Too little residual value. On a GTX 1050Ti worth under RMB 200, a RMB 280 repair makes no sense. Replace it.
- A hard mining history. Cards that ran memory at 95°C and above for months have fatigue everywhere, not in two chips, and the recurrence rate at three to six months is high. We will do the work, but we say so first.
- Blistered or burned PCB. Water damage and burned boards usually mean damage inside the layers, which cannot be repaired.
Guangzhou prices for graphics-card repair: a clean and repaste is RMB 80-150, memory reflow RMB 250-400, a single memory chip replacement including the chip RMB 350-600, and board-level power repair RMB 300-800. ACCPC tests before quoting and charges nothing if you decline the repair after testing, runs a full burn-in before delivery, and warrants reflow work for 90 days.