Share:

Video summary

Luke and GPU Lab’s Greg upgrade two Nvidia RTX 4090s from 24GB to 48GB VRAM each, creating a 96GB local AI setup.

Seeking a more private, self-hosted AI setup, Luke travels to Michigan to have GPU Lab’s Greg upgrade two RTX 4090s from 24GB to 48GB of VRAM each. Greg explains the custom PCB design, which stacks memory modules to reach 48GB, and walks Luke through the delicate BGA work: transferring chips, cleaning solder pads, adding solder balls, and aligning the GPU cores. Luke helps place memory modules, while Greg operates the rework equipment. Both modified cards boot and pass a five-minute memory test, reporting 850GB/s bandwidth within spec; they then undergo water cleaning and a 12-hour oven dry. Greg also describes GPU Lab’s other projects, including 4080 Super upgrades from 16GB to 32GB, A100 conversions, and planned 170HX modifications.

Chapters

  1. 0:00Intro: Luke’s two RTX 4090s are headed for 48GB each and 96GB total VRAM
  2. 2:50What Greg does: GPU Lab upgrades 12–24 cards monthly and uses a 3090-style 48GB design
  3. 3:52Applying memory chips: Greg prepares pads, reflows 12 modules at once, and uses tested salvaged memory
  4. 6:32Good luck, Luke: he places memory, starts reflow, and Greg removes the Dankpods Gigabyte core
  5. 9:55Removing memory modules: Greg uses IR preheating to salvage 12 chips and cautions GPU buyers
  6. 11:21Cleaning memory modules: Greg removes solder and prepares flat, bondable chip pads
  7. 12:32Luke cleans memory pads, rinses flux, and Greg prepares two GPU cores for rebonding
  8. 14:45Adding solder balls to cores and memory: Greg uses stencils, controlled reflow, and ultrasonic cleaning
  9. 17:29Applying memory to the Chinese PCB: Luke places chips for 48GB across both sides
  10. 20:15Attaching the core to GPU 1: optical alignment, reflow, and Luke’s terrifying snap-off
  11. 22:58Attaching the 48GB cooler: GPU 1 passes testing at 850–900GB/s before flux cleaning
  12. 25:54Time to bake: both working 48GB RTX 4090s need 12 hours, while GPU Lab plans 170HX mods

What the full analysis includes

12 chapters · 59 key points · 12 section summaries · mind map

Preview of the first chapter · Intro: Luke’s two RTX 4090s are headed for 48GB each and 96GB total VRAM

  • •Local AI offers a more personal alternative to cloud services0:39
  • •Luke says cloud AI demand is driving up hardware prices0:54
  • •Luke upgrades two RTX 4090s from 24GB to 48GB each1:09

The remaining 56 key points, section summaries and mind map are in the full analysis. Analyze this video to see everything →

This is a Tier 1 public summary

Whether the chapter key points, section summaries and mind map are public is up to the person who shared it. Want the full analysis?Submit one yourself.

More from this channel

Related analyses