All new Registrations are manually reviewed and approved, so a short delay after registration may occur before your account becomes active.
Modded RTX 3080 with 20GB VRAM from China?
As title says.
https://www.reddit.com/r/LocalLLaMA/comments/1ttz558/i_trusted_random_person_on_this_subreddit_and/
https://www.reddit.com/r/LocalLLaMA/comments/1p0bbrl/rtx_3080_20gb_a_comprehensive_review_of_chinese/
I asked around and checked prices. Its roughly 450€ + shipping. My quoted price for 3 fan version was 491€ with shipping.
Apparently people buy these from alibaba directly or via ebay reseller for 600+€
3 fan version
https://www.alibaba.com/product-detail/Good-Quality-RTX-3080-20GB-Graphics_1601612164521.html
https://www.alibaba.com/product-detail/RTX-3080-20GB-Fan-Cooler-GDDR6X_1601732973480.html
Blower/Server version
https://www.alibaba.com/product-detail/RTX-3080-20g-Turbo-Gaming-A1_1601729862967.html
Apparently they just combine VRAM from two original 3080s that ship with 10GB VRAM each.
Has anyone purchased these, or considered purchasing? Would be nice to buy one for AI usage.
Also found V100 with triple fans.
https://www.alibaba.com/product-detail/Customized-V100-SXM2-to-Pcie-Graphics_1601764888923.html

Comments
I remember watching a video from GamersNexus about a shop in China that does this.
Yeah but those are called 4090D models and out of my budget range xD
There are many models, such as 2080ti 22G, 3080 20G, 4090 48G.
You get better performance on a 3090 compared to a 3080.
There is reason why people always buy 3090, 4090 and 5090.
With 30GB VRAM you can run decent models fast.
20GB brings you into to short for these.
3090 has 24 GB of VRAM so not much more.
You could probably fit a 4bit quant into 20GB VRAM.
https://huggingface.co/unsloth/Qwen3.5-27B-MTP-GGUF
One of the best models you can currently run on consumer hardware.
But context its not gonna be huge.
Better would be 6bit quant though.
Look at the benchmarks/money ratio, if it checks out, get it if you can.
just buy a Mac mini like anyone else
My workstation pc can beat the Mac mini lol
But do they run the VRAM at full speed?
Also the 3080 uses a 320-bit bus, the 3090 uses a 384-bit bus. Even if you get nearly as much VRAM on a 3080 as the 3090 has, the memory will be slower because of the smaller bus width.
There are also some 4080 32GBs.
I was wondering who would put time and money into designing this cooler with RGB and stuff but turns out to be some 4090 AIB card. I guess they now have use for all the coolers that were from gaming cards turned into blowers.
I tried 2080ti 22GB before. They run fine but worried about durability due to soldering and modded version. But I got good pcb and reference fan version instead, at least run very cool.
You can buy 2 of them (3080) and run together with nvlink if price is right..
If you are aiming for single card and not going to take advantage of older cards for nvlink, better don't get beyond 3090. They are slow..really slow for AI as single card. 4070ti super / 5070ti 16GB works better for similar price range.
Nothing beats Apple silicon per Watt. Nothing. And you definitely do not have 128GB+ UNIFIED memory.
The bios of the modded card does not support Rebar and the card does not support nvlink.
Other than that, you could distribute the layers or tensor cores across cards with lm studio and the like with minimal impact.
The Mac Mini is great power wise, but doesn't come close to memory bandwidth.
Which is what you want usually.
Unified memory is always a compromise compared to VRAM
It doss support nvlink. Not an issue for 2x2080ti 22gb (44gb).. they r running just fine. The performance is still pretty slow. Token/s isn't appealing but can load pretty large models.
3080 modded cards has heat issues and most likely will be throttled badly on full load on your vram.
Yea but its nice to load big models actually.
It might not be fast but it runs.
With same logic, just get 128GB RAM kimsufi server and run it on pure CPU. I see no difference in 5 tokens per second and 10 tokens per second.
No, Mac Mini is gonna be way faster than a Kimsufi on CPU.
Gonna depend on the model though, obviously a 400B model isn't gonna be fast but a 122B non dense would fly on the Mac.
On the Kimsufi it would be considered slow.
They simply swap the RAM for double sized chips, this works with retail BIOS and in this case engineerig sample ones. Speed is the same.
I doubt any support REBAR but i see no reason why no NVLINK.
Yah, They r same cards with more vram.. nvlink support and all others are almost identical as the base card.
REBAR relies on knowing how muchh memory the card has so i see why this could fail but yes NVLINK should work just fine.
These are old mining cards now with new memory chips - as crypto is very memory intensive (mostly) you actually get "better" quality with the new chips.
The new ram is likely sourced from dead mining GPUs (any brand, same chips) or uust bought at retail.
The DIE will work many years more even if the card was treated horrible, the new memory if properly cooled (a known Nvidia problem) gives it some yeares more life.
Im not into the cost side of things as my job uses grey market AI cards imported to China rather than consumer cards/mods.
Above ~160€ Aliexpress does not apply VAT to EU customers, VAT will be applied on import plus import duties and exaction tax. 491€ becomes 687€ which is coarsely +40%.
Time for china vacation
Order also sexbot or sexdoll with it, atleast you would get something in return if the card is fake..
and what use case would i need that for? Just Curious. Are you just going to run LLM all day? lol
That is the wisest advice I found in this thread, by far.
I was fearful I had been wasting my time reading this on a Saturday afternoon but you saved me, thanks.
yes, that is the use case, inference also
the TB5 links between the macs are also very useful
With two suitcases packed with video cards worth 50,000€ it will be a long day at the airport interrogation room.
Just say you're adding more ram on GPU for the kids in your neighborhood so they can play more vidya