AMD RADEON PRO V620 32GB GDDR6 GPUs - $565
- Thread starter jbrukardt
- Start date
Notice: Page may contain affiliate links for which we may earn a small commission through services like Amazon Affiliates or Skimlinks.
I just went on alibaba.com, searched for AMD Mi50, and messaged one of the sellers with a lot of sales. PM me if you really want my specific seller.
Can't guarantee you'll get them for $150 in the US, what with how fluid your situation is with taxes. I ordered a dozen a few days ago for $141 delivered to Germany with DDP shipping.
The MI50s shipped to the US are more like $230-250 each after fees (Alibaba) and tariffs. I just ordered 4 of these myself, the top seller (first result for "Mi50 32gb) ships via UPS and you can pay tariffs through there if you choose the non CDP shipping listing. They arrived in 3 days after shipping (INSANE!!!).
yes, someone has, just found this today: https://github.com/PowerfulGhost/vllm-mi50
my fav seller didn't want to go down with the price as I wanted so still unsure what I'll do.
searching some more the community support seems to finally pick up and quickly outpace AMD
- GitHub - nlzy/vllm-gfx906: vLLM for AMD gfx906 GPUs, e.g. Radeon VII / MI50 / MI60 ("vLLM for AMD gfx906 GPUs, e.g. Radeon VII / MI50 / MI60")
- GitHub - mixa3607/ML-gfx906: ML software (llama.cpp, ComfyUI, vLLM) builds for AMD gfx906 GPUs, e.g. Radeon VII / MI50 / MI60 ("ML software (llama.cpp, ComfyUI, vLLM) builds for AMD gfx906 GPUs, e.g. Radeon VII / MI50 / MI60)
- GitHub - ai-infos/guidances-setup-8-mi50-llm: Guidances for Test setup of 8 AMD MI50 32GB (LLM) ("Guidances for Test setup of 8 AMD MI50 32GB (LLM)")
- GitHub - swiftfr13/MI50-gfx906-Infos: Guide on running MI50 as a normal GPU in Windows ("Guide on running MI50 as a normal GPU in Windows")
All AMD lobbying/nagging I do myself is for the Radeon VII Pro / gfx906
My workhorse is a dual 2667v4 which can theoretically power two MI50 though I would stick with one. Price wise it's all in favor of the MI50 especially if the ecosystem now picks up and it's no longer such a brain cancer trigger to make it work with ROCm (at least the debian builds have it enabled in ROCm 7, though all warnings apply, after all AMD still isn't putting it back on "working but unsupported" status.
I feel I should add a disclaimer, or warning: you should really think about if you want to use something where the vendor does everything in their power to fight you being able to use it. I know they're improving but - so far - the actions are not yet making the difference that would make the warning obsolete. I know if I had gone with a much, much more expensive NVidia I would have had more time to do something, and if I would go with some utterly unsupported Tenstorrent I would at least learn everything from the basics upwards.
Last edited:
I'd also include this reddit thread. It's the same guy who wrote the P40 kernels for flash attention back in the day. While it may not be as fast as vLLM, I personally prefer llama.cpp because of wider GGUF model availability and quant support, and ease of running/switching models. Coupled with llama-swap, it is a pretty powerful combo for loading or switching between multiple models.searching some more the community support seems to finally pick up and quickly outpace AMD
I know this was about the V620 initially but the current conclusion for _that_ hasn't changed.
- GitHub - nlzy/vllm-gfx906: vLLM for AMD gfx906 GPUs, e.g. Radeon VII / MI50 / MI60 ("vLLM for AMD gfx906 GPUs, e.g. Radeon VII / MI50 / MI60")
- GitHub - mixa3607/ML-gfx906: ML software (llama.cpp, ComfyUI, vLLM) builds for AMD gfx906 GPUs, e.g. Radeon VII / MI50 / MI60 ("ML software (llama.cpp, ComfyUI, vLLM) builds for AMD gfx906 GPUs, e.g. Radeon VII / MI50 / MI60)
- GitHub - ai-infos/guidances-setup-8-mi50-llm: Guidances for Test setup of 8 AMD MI50 32GB (LLM) ("Guidances for Test setup of 8 AMD MI50 32GB (LLM)")
- GitHub - swiftfr13/MI50-gfx906-Infos: Guide on running MI50 as a normal GPU in Windows ("Guide on running MI50 as a normal GPU in Windows")
All AMD lobbying/nagging I do myself is for the Radeon VII Pro / gfx906
A caveat with llama.cpp on gfx906 is that -sm row still doesn't work for most models, but I don't mind because I use them to load larger MoE models anyways.
If you have ~350W to spare on your PSU, you can install two of them comfortably. I limit mine to 170W and haven't noticed any degredation in performance. They idle at 16-22W, whether there's a model loaded or not.
Setting up ROCm was a 15 minute affair on my first system, and that was before all the tutorials. The issue with the TensorFiles was solved with a google search and opening the first result. The rest was following AMD's own instructions.
AMD's support has always been less than stellar. Look at the driver fiasco over the weekend with RDNA1/RDNA2. But in the case of the Mi50, they work in Linux out of the box without doing anything.
At the end of the day, it's a tradeoff. In the current market, if you want support, you need to spend several thousands per card because we're in an AI bubble. Heck, ECC DDR4 is becoming unobtaainium now.
I personally think the Mi50s are great pieces of kit for the price (I got mine for 135 a piece). They offer the same VRAM of the 5090 at 4090 VRAM bandwidth. Compute is better than the P40 at FP32, and much much much much better than the P40 at FP16. There's nothing in the market that comes even close. I built a 192GB VRAM rig for ~€1.6k.
That's not a fair comparison. The Mi50 supports Vulkan on linux out of the box, AFAIK. So, it was always functional, if not optimized for.
As much as I'm rooting for Jim Keller, Tenstorrent requires you to write your own inference code. If you can do that, I'm fairly sure you'll find a job at Tenstorrent themselves.
Hi Folks! Question.. for those who have MI50's and V620's.. do the rear section have the same screws alignment?
Or is there a difference?
I'm trying to get a blower adapter but unsure why the V620's adapter offers cost 20% more and have shipping cost on top vs the MI50's adapters.
Or is there a difference?
I'm trying to get a blower adapter but unsure why the V620's adapter offers cost 20% more and have shipping cost on top vs the MI50's adapters.
I have lots of Mi50s and I keep them cool without neither breaking the bank nor sounding like a jet engine.
If you have any rudimentary 3D modeling skills, you can design a 3D printable shroud that fits the rear of the cards and lets you cool each pair of cards with a single 80mm fan.
If you lack said modeling skills, you can still make such a duct from thin plywood or PVC sheet.
The cards are about 97mm high. A single PCIe slot is 20.32mm wide. So two double width cards are conveniently 81.28mm. That's an 80mm fan, with the remaining 17mm for Pasing the power cables out.
I use Arcitic S8038-7k fans to cool each pair.
Last edited:
Unfortunately no 3d printer and I so far have had no response on my local business that 3d print.
Also I currently have only one card. Hence why I was thinking in the snail fan approach.
Might try to make a rudimentary tunnel using cardboard + heavy duty packing foam + gorilla silver tape. The question would be how much air pressure does the M620 needs to cool properly.The fins seem pretty dense.
It seems that nobody is really reading along the lines of what I wrote.
The question was about COMPATIBILITY. Not if there is availability or not. I know there is availability.
Nor about 3d printing your own stuff or design your own stuff ( If i had a printer I would have done so already or download a print file available).
As for EBAY, I only seen ones with 2 micro fans, server style.
Or with a 60mm fans. Noone with snail style blowers that mention compatibility with the V620, only MI50. (again, hence the question about compatibility and hoping they have similar rear design).
One thing that I noticed is all MI100, MI50 and older V340 and 540 having compatible rear spacing and screws. But no mention on V620's.
Kinda weird all the talk about MI50's when the thread is supposed to be titled after the V620.
As for my point, the issue is COST.
Any aliexpress product that offers actual V620 compatibility costs 30% more than a similarly looking MI50 one.
With MI50's even offering free shipping, making the total cost of one vs the other around 40% more.
Not that much from my experience. I also had your same thought when I looked at the heatsink the first time, but in reality they're no worse than a CPU heatsink.
Before building my rig, I had two in a Xeon board on my desk. I grabbed a regular 5k rpm 80mm fan and some double sided tape, and stuck it to the metal brackets as an exhaust pulling air through the cards. I used an aliexpress 3pin to molex cable jerry rigged to fit four pins to run it at 5v instead of 12v because it was too loud. That was mostly enough to keep both cards cool running dense models for a minute or so at a time (as in, active inference time) before getting to 80C. Letting the fan run on 12v and the full 5k, temps were under 70c for the same test.
There are lots of fan duct designs on sites like thingiverse and printables. JLC is cheap if shipping is cheap to your country (for europe, they offer EuroPakt). An 80mm fan will do fine, but if you care about noise, then a 120mm Arctic P12 Pro/Max will me more than enough for a single card.
There's no need to be rude to people who are trying to help.
If you feel everyone doesn't understand what you're saying, maybe consider that you haven't explained your issue clearly?
is 5-15 usd too expensive?
there are all kinds including full shroud replacement.
ex. Radeon Instinct MI25, MI50, MI60 Full Shroud Replacement | eBay
the limit is your imagination... Those cards weren't meant for desktop; but workstation or server.
If you want desktop solution go with normal desktop card instead.
or better yet create a thread for that specific case you are looking for - not the v620 that amd closed the conversation about ever releasing proper drivers for.
I am unsure which part was rude to you.
I literally was trying to explain the issue in a better way by being specific.
And it is kinda weird considering you just said "maybe consider that you haven´t explained your issue correctly".
Is my point of asking if the MI50 rear section has the same size and same screw positions as the V620 to install a snail blower type fan really that confusing?
Noone of them are 5-15 USD?.
I see 20+ USD plus Shipping (which is 20 USD on average on ebay). Note I am in Canada.
AliExpress adapter + snail fan is around 40-45 CAD for the part only and then 15 USD to 20 USD shipping.
Mi50 are 30 CAD on average and free shipping (again, hence my question about compatibility and just buy a Mi50 and plop it on the rear of mi V620.
Thanks! Will ask the shop vendor if they are compatible with the V620 (As I remember reading that they have different mounting points). Unless of course I decide to risk it by using tape XD
[/QUOTE]
the limit is your imagination... Those cards weren't meant for desktop; but workstation or server.
If you want desktop solution go with normal desktop card instead.
or better yet create a thread for that specific case you are looking for - not the v620 that amd closed the conversation about ever releasing proper drivers for.
[/QUOTE]
"The limit is your imagination" . ID say cost is it right now.
"If you want desktop solution go with normal desktop card instead." The price difference of almost 400 USD is definitively a factor vs a W6800 PRO and a desktop solution would have had 16GB or less of vram.
"or better yet create a thread for that specific case you are looking for" I am confused. This thread is literally about the V620. Why should I create another thread ?
"note the v620 that amd closed the conversation about ever releasing proper drivers for." I am running linux atm.
*edit*
Forgot to ask: You guys would recommend using a full shroud replacement vs an external fan with adapter?
*edit 2* I am an idiot, I just noticed that the parent forum category is DEALS, as in finding cheap prices on the card itself, not talking in details of the card itself.
You can ask him to make it compatible with v620, after all he's 3d printing those.
Create your own thread, and we can go from there.
Did, now to wait for the answer
In the other hand, as for new thread: https://forums.servethehome.com/index.php?threads/amd-v620-shroud-fans-options.54660/
Anyone try the drivers from Azure for their V620 VMs?
Azure N-series AMD GPU driver setup for Windows - Azure Virtual Machines
How to set up AMD GPU drivers for N-series VMs running Windows Server or Windows in Azure
learn.microsoft.com