[DEAD] Ebay: Used XFX BC-160 pulls, $30 w/FS

johnny0

Weaksauce
2FA
Joined
Jan 8, 2017
Messages
122
Seller: networktradersss (https://www.ebay.com/usr/networktradersss)
Link: https://www.ebay.com/itm/398348164999, https://www.ebay.com/itm/398359058789

Looks like a bunch of passive BC-160 v2.0 cards have been decommissioned and are being liquidated. The pictures show a clean pull, but expect that these have been crunching 24/7 for years. $30 shipped.

Navi 12 with 8GB of HBM2. No display outputs, VCN block is likely disabled. These need special consideration for cooling, know what you are getting into.
 
Last edited:
As an eBay Associate, HardForum may earn from qualifying purchases.
What "can be done" with these? Mining? LLM? or too old and slow ? ( looks to need 2U chassis and front to back high air flow to have any luck at cooling these or a rigged up hack solution )
 
A while back I remember seeing people use two cards for frame generation, so maybe that for gaming?
 
( looks to need 2U chassis and front to back high air flow to have any luck at cooling these or a rigged up hack solution )
Use a CPU Cooler for the GPU, tighten with zip ties, put in a cable organizer box with a 120mm fan for airflow in, and another 120mm fan for airflow out, and a PCIe riser cable. If it's short enough, and you have enough PSU cables, you can use your PSU to power it, or just use a secondary PSU to power it.
 
$120 for 4x rx5700s spread across llama for cheap 35b model processing? Could be very cool
I was wondering how well this would work too, new to running local ai models but isn't nvidia much more desirable, will most models run on amd gpu's?
 
I was wondering how well this would work too, new to running local ai models but isn't nvidia much more desirable, will most models run on amd gpu's?
Most utilize CUDA, but maybe one could use an NVIDIA card to run on the model, while utilizing the AMD graphics card memory? Maybe it's not there, yet. As an aside, ROCm will likely get more popular as NVIDIA card prices skyrocket.
 
Most utilize CUDA, but maybe one could use an NVIDIA card to run on the model, while utilizing the AMD graphics card memory? Maybe it's not there, yet. As an aside, ROCm will likely get more popular as NVIDIA card prices skyrocket.
If rocm supported these they would be like 500 dollars each probably lol
 
One guy on Reddit got it to work, might be worth the experiment
 
Could use them for Moonlight/Apollo boxes to stream in a gaming cafe or if you wanted to have some Minecraft VMs using OpenCL mods to increase world generation.

Not much use for most of us.
 
I missed out on a complete x299 rig locally for $500. Could have supported 4 no problem.

I currently have an x570 godlike that runs a 2080S+1080Ti+2x1080 in a Frankenstein LLM rig with no issues. Using a 35b abliterated qwen model, no issues on Windows. Not sure how it would do on the AMD side
 
OOS. I received a clean v1.0 card. OP updated.

The last restock/relisting was quick, will give it a couple days before marking dead.

Update: Relisted at $45. RIP.
 
Last edited:
OOS. I received a clean v1.0 card. OP updated.

The last restock/relisting was quick, will give it a couple days before marking dead.

Update: Relisted at $45. RIP.

lol typical though.. you posted it... it starts selling.. demand goes up.. Seller decides to see how much more they can squeeze out of them
 
Back
Top