https://www.ebay.com/itm/177571378959
Chips are starting to pile up on eBay and the price is less than most R9700s…
https://www.amd.com/en/products/accelerators/instinct/mi200/mi250x.html
128 GB of HBM2E
3200 GB/s
Before anyone gets too excited, the plug is easy, it’s the power delivery and BOIS that are the blockers. If the BIOS can be solved, then it’s just power delivery.
I’m hoping someone overseas is working on it…
The UBB system is tough, the MI350P looks like a PCIe card, but again the power servers like this use is outrageous. Handful of used systems and boards you can reverse engineer? The SXM / OAM is unique and high power/high thermal transfer is assumed.
600w(MI350P)*8 modules 4800w board, 22A 220v, or 40A 120V to power a single 8 GPU module. One is ~500w(MI250) isn’t far behind and would need it’s own PSU balanced to your system.
Yeah, it’s a tough problem to solve, but not impossible. The ROI just needs to be there. If the GPUs start dipping under $1k then it’s likely not long before someone overseas decides YOLO and makes a PCIe baseboard with 3-4 8Pin (please let the 12pin die…).
For anyone else then me who does not know the specs of these cards.
MI250X
128 GB of HBM2E
3200 GB/s
The price comparison between R9700 made me expect a much less potent card. I hope these chips gets turned in to pcie GPUs
Yeah, they are already a good ROI at ~1400, but under 1k is where it gets VERY interesting.
Relevant front page of HN post today “I Put a Datacenter GPU in My Gaming PC for £200 :: The Tymscar Blog”
if only it was actually that affordable, I’d have more than one.
POC’ing the 16GB for my stack, then going to swap for 32GB variants.
The MI250s are the dream though.
Yeah ~600 entry(watercooling is awesome!) Moving to 32GB would be my first step. PCIe cards are ~750 plugnplay. Expansion is a feature though
Here to see the build though
We need a dedicate thread to follow along with this build.
It has gone through a few phases. Started out as a 5700G with 2x 7900XTs.
Then swapped it out for a single R9700.
Then 2x R9700…
Then upgraded to a 5950X…
Then a 3960X with AsRock Creator board.
Then started playing around with PCIe Switches to get around the P2P issue with AMD consumer cards.
The PCIe switch card got REALLY hot without server airflow. So got creative.
Then needed a way to organize the mess.
Also put 2 pumps in series for better flow and redundant cooling.
Probably good for compute and AI work, and if you are doing that, why not just use them as intended instead of trying to adapt them them into a PCIe slot?
They are probably terrible for gaming, as the drivers are likely not optimized for that, and likely miss many key features.
I don’t think anyone here is going to use them for gaming.
The reason for PCIe is not everyone wants to run them in a baseboard (a single card with 128GB of vRAM would be more than enough for most people who want a local LLM setup), and it paves the way for setups like this where you can adapt the units to a PCIe switch vs needing a custom motherboard that has the connectors needed.
128GB for 4300 sketchy dollars
Also a gamble. But Something I’m working towards with the 2x16GBs as a POC. If the NVlink works between the 2x then the 4x board will be ordered and tested with two more 16GB before diving into the 4x 32GB. I’ll likely have to sell my 2X R9700s to fund that, but for less than the cost of a retail spark, you have something that will run some SERIOUS models. Add another 4x board and you have replaced most Cloud services.
This is where the MI250s are very enticing… a 2x board is 256GB, a 4x board would be 512. That’s Frontier level compute in your home for less than ONE RTX 6000.
I was looking for what AIO cooling options there is for SXM2 cards and came across this eGPU set up. 4x32gb v100 in one chassi bundled with PLX pcie switch. I think premade solutions like this will make a lot of sense for some users. I would hope MI250 gpus get the same treatment. This would solve power and noise.
The V100 is SXM.
The Mi250 is OAM. Very different standard.
I think they were more sharing that in the past the SXM standard was also something that was once an enterprise server option only. Then one day V100s got cheap enough that someone put in the effort to adapt them to PCIe cards. Now it’s trivial to toss them into a gaming system to run local models.
OAM is indeed a different standard. But in some ways it’s a better one since it’s more open-source. Once someone puts the effort into figure out the power and bootstrap blockers, there is little to no red-tape to open the floodgates on OAM boards.
