$142 upgrade kit and spare modules turn Nvidia RTX 4090 24GB to 48GB AI card
Posted by cornucopea@reddit | LocalLLaMA | View on Reddit | 51 comments
The upgrade kit comprises a custom PCB designed with a clamshell configuration, facilitating the installation of twice the number of memory chips. Most components are pre-installed at the manufacturing facility, requiring the user to solder the GPU and memory chips onto the PCB. Additionally, the upgrade kit includes a blower-style cooling solution, designed for integration with workstation and server configurations that utilize multi-GPU architectures.
CommunicationStrict@reddit
Where do you guys get the upgrade kit from?
CertainlyBright@reddit
$142 is a barebones price for high volume, almost raw cost for the board. Good luck finding it for that price. You also need 12 more D8BZC modules which add another few hundred dollars. And you are not taking off a core or the memory modules without specialized equipment that costs thousands of dollars and requires lots of experience- At least without overheating all components and taking years off the life of the vram. Theres a reason 48gb 4090's cost 1.5x the cost of an aftermarket 4090
az226@reddit
I’ve been offered the PCB for $150. Single units.
ZeroGee0@reddit
I must be doing it wrong. Can't find much close to that. Mind sharing? I've got most of the equipment and previous experience...
az226@reddit
Where are you located?
AcePilot01@reddit
usa?
Leopold_Boom@reddit
Any idea why there isn't a 3090 version yet?
az226@reddit
Because it’s the past. There is more value in the 4090 being upgraded than the 3090 being upgraded.
Simply, follow the money.
The real question is why they make the PCBs from such low quality materials when they can increase the cost value and reliability of the card if they invest $50-100 more.
Swashybuckz@reddit
Why didn't Nvidia just make the fucking thing 48megs. They clearly could have.
az226@reddit
5090 was supposed to be 48gb but in the final few months they decided to gimp it to 32gb.
baldden@reddit
Where is you found the pcb? Link? Thank you
ZeroGee0@reddit
I must be doing it wrong. Can't find much close to that. Mind sharing? I've got most of the equipment and previous experience...
a_beautiful_rhind@reddit
You can do the memory chips at home. The core tho.. nah.
Operating temp is not the same as heating temp. If the chips degraded at soldering temps, there would be no way to attach them in the first place.
Wrong-Historian@reddit
Yes, they can stand the soldering temps for a few seconds. The solding temperature diagram is in the spec sheet. So during manufacture, that's computer controlled.
Hence the '1000's of dollars of equipement required'.
Sure, you might reflow it with some hot air, but you'll either have to use much higher temperature and/or for much longer time than a machine can do it, to get a reliable reflow.
a_beautiful_rhind@reddit
Repair guys get it done. Videos of how the cards are made, as posted here, show them just using a stationary hot air. You don't need a pick and place/computer controlled reflow.
RightToBearHairyArms@reddit
Having reballed Xbox360s, you’re definitely right that this is outside of 98% of peoples ability. And the 2% that can do it likely didn’t have a perfect success the first try.
serp7777@reddit
Sorry voor de off-topic vraag. Kan iemand een werkplaats in Nederland aanbevelen waar ik het geheugen van mijn videokaart/moederbord kan laten repareren of opnieuw kan laten solderen? Ik vermoed dat de geheugenchip is losgekomen aub?
prusswan@reddit
Really goes to show how VRAM is dirt cheap, it is just the technology that is expensive
tinny66666@reddit
Right? It would just take one mediocre player to shower us with VRAM and they'd become an instant hit, get software support and become a real player.
datbackup@reddit
poor software support. AMD in particular has been notoriously bad about bugs and jank in the ROCM stack. It has gotten a lot better in the last year though but the leader is CUDA no question.
sunk costs. Big deployments have people on staff who know CUDA/nvidia and their quirks. Not to mention the existing code bases that target CUDA and would take lots of time, effort and money to rewrite targeting another framework.
urgency to become and remain competitive. The game is to deploy AI ASAP in order to be not be left behind as the torrent of advances continues and intensifies. Choosing an unproven framework is an excellent way to waste lots of time and money only to eventually spend even more time and money when it’s decided to go with the field leader.
I do think we will see a big push from someone. It’s weird that google just keeps all its tensor hardware for itself and doesn’t try to sell or license it. Shows what exactly is on the big players’ minds—-winner take all.
tryingtolearn_1234@reddit
Intel especially. Unfortunately what will happen instead is China will use this obvious gap to break in to the market.
repolevedd@reddit
In the text, the author somehow distorted the meaning and for some reason came up with 'hacked firmware.' In the video, the original author carefully explains that modifying the VBIOS of 24 GB graphics cards to support 48 GB is impossible due to the Falcon chip in the GPU, which verifies checksums, so a 48 GB VBIOS could only have been created by Nvidia, which has the certificates to sign such firmware.
jakegh@reddit
Sure, all you gotta do is reball an AD102, a 600mm2 chip couldn't have more than what, 2-3 thousand balls? Nothin' to it!
FullstackSensei@reddit
You only need to solder the GPU and memory chips!!!
Makes it sound like modding an old Trident VGA card!
SkyFeistyLlama8@reddit
Trident were good cards before 3dfx Voodoo came and made 3D accelerators mandatory for new PCs.
FullstackSensei@reddit
Having grown up with a CGA monitor, Voodoo cards really felt like voodoo when they came out. 3D games were never the same again.
rocket_randall@reddit
I still have my Quantum Obsidian X24 around here somewhere
SkyFeistyLlama8@reddit
We found the oldie LOL.
I remember seeing VGA 256 color screens looking like magic after CGA, and then 16.8 million colors looked like a vision of paradise, and now I'm staring at a high density OLED screen with sweet DCI-P3 coverage. Next up is a neural interface. For work, of course.
lemondrops9@reddit
the glory days of 3D. I still have my voodoorush card.
Cacoda1mon@reddit
Just get a Matrox Mystique and add the 2 MB memory upgrade. No soldering required.
m1013828@reddit
imagine if they just had the onboard memory plus some ddr5 slots on graphics cards. complexity of another ram/cache tier but stops some of the e wastage getting more lifr out of cards
Long_comment_san@reddit
Why ddr5, make it a new memory form factor. Why not slap some LPDDR5X chips on the backplate?
m1013828@reddit
just thinkijg from cost, optional dimms or sodimms as they are an existing format. cant trust nvidia to do a proprietary format.....
a_beautiful_rhind@reddit
Some bright upstart gotta start a service. Double your memory for $500 or something.
prusswan@reddit
I hope their workmanship is at least equal to Brother Zhang:
https://www.youtube.com/watch?v=1H3xQaf7BFI&t=9292s
Vegetable_Low2907@reddit
Link? I don't think most people realize how difficult it is to reflow memory chips without damaging hundreds of other components on boards as jam packed as the 4090!
a_beautiful_rhind@reddit
I see plenty of people fixing GPUs and replacing mosfets or memory on youtube. All they have is a heat gun and maybe a board warmer. I agree with you that it's a skill though.
segmond@reddit
This is localLlaMa
Vegetable_Low2907@reddit
I see the Tom's Hardware link now :)
shroddy@reddit
We provide you with the circle, you only need to draw the rest of the wolf.
Scolder@reddit
This article covers it in more detail - https://videocardz.com/newz/modder-turns-geforce-rtx-4090-24gb-into-48gb-card
segmond@reddit
Picture? Links?
sautdepage@reddit
It's one of the best part of Gamers' Nexus recent documentary. Awesome but clear that most people shouldn't attempt this at home.
https://www.youtube.com/watch?v=1H3xQaf7BFI&t=2105s
CentauriWulf@reddit
Yes please. Without this info, this is a very difficult post to care about.
Impossible-Glass-487@reddit
Agreed. WTF?
__JockY__@reddit
Typically one would cite a reference, provide a link, or give some other information that would provide the reader options for further detail. Your post lacks all of those things, so we’re left with nothing but “cool story, bro.”
cornucopea@reddit (OP)
Link is added, it's my mistaking how to create a new post, one of tab says "Link" so I moved the link to there, and didn't bother check after submission until I came back later saw the complains.
Mediocre-Method782@reddit
You dropped this
https://www.tomshardware.com/pc-components/gpus/usd142-upgrade-kit-and-spare-modules-turn-nvidia-rtx-4090-24gb-to-48gb-ai-card-technician-explains-how-chinese-factories-turn-gaming-flagships-into-highly-desirable-ai-gpus
cornucopea@reddit (OP)
I put it in the "Link" box when created the post. I should have just included the Url at the bottom of the text.
Jekyll0101@reddit
chinese sellers are charging around 5000 rmb for solely 4090 24g vram to 48g vram upgrade service , which is around 700 usd
hitmanactual121@reddit
OP looks like they are referencing this post from Toms hardware: https://www.tomshardware.com/pc-components/gpus/usd142-upgrade-kit-and-spare-modules-turn-nvidia-rtx-4090-24gb-to-48gb-ai-card-technician-explains-how-chinese-factories-turn-gaming-flagships-into-highly-desirable-ai-gpus