Been trying to find a proper fan/blower for it.
Most people seem to use the classic adapter+snail blower at the end of the card.
Others recommend a custom 3d printer shroud.
Problems: AliExpress V620 snail cooler blowers(which look almost identical to Mi50 ones) cost approximately 40% more than the MI50 ones and they want to slap shipping on top (the mi50's are free shipping).
No vendor knows if there is compatibility ( I assume they just print the 3d models someone else created for the Mi50 and call it a day after adding the blower).
Note that I lack of 3d printer and the local stores have not answered my quote requests.
So.. questions:
Are the rear (where the power connectors are) screws similar between the Mi50 and the V620? (from a visible standpoint they seem very similar but then I do not own a mi50).
Would it be better to use a custom full shroud than a snail blower?
My budget was kind of blown right now. So I want to keep costs low. Or should I just bite the bullet and buy the overpriced V620 versions thru AliExpress?
Most server gpu's have screws in same positions (amd and nv) thus all 3d printed snail blowers will work on all datacenter gpu's as they are shipped for server/workstation standard. *only differences in most nv cards you don't have the top screw - you only have 3 on the bottom. Size wise they are also the same.
The mentioned full shroud replacement will likely need the 3d model adjusted.
Most of those snail shrouds, and full shrouds do not include fan - keep that in mind. You will have to spend more.
Alternative option is to buy high pressure fan, and mount it outside in pull configuration like supermicro workstation do. (its also what i use for passive cards.)
2. What is better?
custom full shroud gives you more space in your case *depending obviously what solution you choose.
all solutions will most likely be loud.
Sorry what about the $20 ones on ebay? I don't know if they work well or not, but 20 seems reasonable.
My local library offers access to a 3d printer, any chance yours does? If you had a design already made, I could also print you one and snail mail it for the price of shipping (I have my own printer).
Sorry what about the $20 ones on ebay? I don't know if they work well or not, but 20 seems reasonable.
My local library offers access to a 3d printer, any chance yours does? If you had a design already made, I could also print you one and snail mail it for the price of shipping (I have my own printer).
Jesus, finally got a response from one of the local 3d printing companies.. They want an insane 100 CAD for a full shroud with no adjusts and 60 CAD for the snail fan style lol.
Pretty sure the MI50 shroud is incompatible. The snail style rear fan might be. I mean the size is the same but the difference is the screws I believe.
I assume it might be usable if you use a wider hole on the horizontal space.
Anyway I did bit the bullet and I ended buying one from AliExpress supposedly for MI620.
Well see how it goes.
Are you using it for LLMs? If you are, do you mind sharing how the V620 is performing for you? I'm debating getting one for an external GPU for my LLM server - thanks!
Are you using it for LLMs? If you are, do you mind sharing how the V620 is performing for you? I'm debating getting one for an external GPU for my LLM server - thanks!
Not to answer for the OP, but I've had a Radeon Pro V260 for quite a while now and it mostly collects dust. I hook it up from time to time for testing purposes, but that is about it.
The first hurdle is cooling. It's a datacenter card made to be used in a chassis with lots of front-to-back airflow. I've used different cooling solutions until ultimately settling on what looks like the same 3D printed fan shroud used by @cesarinpillin. It's the most effective at cooling the card while being the least obnoxious, but it is still loud under full load. The blower style fan makes a less annoying sound than any of the square fans I've used which had adequate CFM needed to cool the card (ranging from 40mm to 92mm). The plans I used are from AMD Instinct MI50 blower fan adapter V2 by MagentL and the fan is a EFH-08E12W-JP01 sourced from Aliexpress. The form factor of this shroud is also better for leaving room to either side of the card, while only being a few milimeters longer than the fan shroud I used for a 92mm square fan.
Second hurdle is getting the amdgpu driver to actually recognize the card. Depending on the motherboard used I've had to enable a combination of driver options on the kernel command line in order for it to work correctly. I want to say I had at least one Supermicro motherboard that I never could get the GPU to work on, but that may have been prior to me uncovering the driver options needed.
Depending on what you want to do it may suffice. Token generation speeds are decent. Infill/Prompt processing speeds are nowhere near even an entry level RTX 50xx card. This makes a difference if you're using this for anything with large contexts (agentic coding, for example). I've included some simple inference benchmarks at the end of this post if you're interested. I've also found it to be much slower at image generation than CUDA cards using ComfyUI, but I don't have comparison numbers for that at the moment.
These are a (relatively) inexpensive way to get access to decent amounts of VRAM. I don't regret buying it from a tinkering perspective. From an actual daily use perspective it just isn't worth it to me. Plenty of people still use MI50/60s, so maybe I'm in the minority. The V620 will be faster at prompt processing than the MI50/60, but a little slower at token generation (faster compute, slower memory).
llama-bench benchmarks for a small dense model and a small MoE model:
Not to answer for the OP, but I've had a Radeon Pro V260 for quite a while now and it mostly collects dust. I hook it up from time to time for testing purposes, but that is about it.
The first hurdle is cooling. It's a datacenter card made to be used in a chassis with lots of front-to-back airflow. I've used different cooling solutions until ultimately settling on what looks like the same 3D printed fan shroud used by @cesarinpillin. It's the most effective at cooling the card while being the least obnoxious, but it is still loud under full load. The blower style fan makes a less annoying sound than any of the square fans I've used which had adequate CFM needed to cool the card (ranging from 40mm to 92mm). The plans I used are from AMD Instinct MI50 blower fan adapter V2 by MagentL and the fan is a EFH-08E12W-JP01 sourced from Aliexpress. The form factor of this shroud is also better for leaving room to either side of the card, while only being a few milimeters longer than the fan shroud I used for a 92mm square fan.
Second hurdle is getting the amdgpu driver to actually recognize the card. Depending on the motherboard used I've had to enable a combination of driver options on the kernel command line in order for it to work correctly. I want to say I had at least one Supermicro motherboard that I never could get the GPU to work on, but that may have been prior to me uncovering the driver options needed.
Depending on what you want to do it may suffice. Token generation speeds are decent. Infill/Prompt processing speeds are nowhere near even an entry level RTX 50xx card. This makes a difference if you're using this for anything with large contexts (agentic coding, for example). I've included some simple inference benchmarks at the end of this post if you're interested. I've also found it to be much slower at image generation than CUDA cards using ComfyUI, but I don't have comparison numbers for that at the moment.
These are a (relatively) inexpensive way to get access to decent amounts of VRAM. I don't regret buying it from a tinkering perspective. From an actual daily use perspective it just isn't worth it to me. Plenty of people still use MI50/60s, so maybe I'm in the minority. The V620 will be faster at prompt processing than the MI50/60, but a little slower at token generation (faster compute, slower memory).
llama-bench benchmarks for a small dense model and a small MoE model:
Running Fedora Server 42, 43, and Rawhide on a few different Asrock and Supermicro motherboards I've just gotten in the habit of adding the following to the kernel boot options:
It's possible other distributions set those options out of the box, or maybe they aren't needed for your hardware.
I'm not sure what you mean by "How long had you tested?" The length of testing would not have any bearing on the card being recognized. At least on a few of my motherboards without the correct driver options amdgpu wouldn't load correctly and the GPU wouldn't be recognized by tools such as rocm-smi, vulkaninfo, etc.
Running Fedora Server 42, 43, and Rawhide on a few different Asrock and Supermicro motherboards I've just gotten in the habit of adding the following to the kernel boot options:
I'm not sure what you mean by "How long had you tested?" The length of testing would not have any bearing on the card being recognized. At least on a few of my motherboards without the correct driver options amdgpu wouldn't load correctly and the GPU wouldn't be recognized by tools such as rocm-smi, vulkaninfo, etc.
I mean as in how long has been since you tested your video card, of course "testing" as in performing specifically testing is irrelevant.
I mean Testing as in owning the device and using it under various configurations.
I've had the V620 for just over a year. I used it regularly for a few months when I first got it, then I replaced it. Now it only comes out if I need some extra VRAM (rarely the case these days) or if I want to see how Vulkan performance is progressing.
A lot has changed from then to now. The Vulkan backend in llama.cpp performs considerably better now. Local models have made massive strides.
I'm not trying to crap on the choice of a V620 Pro. In fact, I think it's an overlooked option for a lot of people. It just has limitations.
> The V620 will be faster at prompt processing than the MI50/60, but a little slower at token generation (faster compute, slower memory).
has anyone tried one of each in the same system, pushing pp to the v620 and tg to the mi60? i have an epyc supermicro board sitting around, and it might be fun to try to have claude optimize the relationship between the two?
For your V620 cooling situation, the easiest low-cost approach is usually sticking with a blower-style fan, since the fan bearings Supplier and airflow path are designed specifically to handle the GPU’s heat and pressure without creating vibration or premature wear. The rear screws near the power connectors on the V620 do look similar to the Mi50, but even slight differences in spacing or thread depth can affect how the blower mounts and whether the bearing-loaded fan spins freely. A custom full shroud can work, but it often requires precise alignment to avoid stressing the blower bearings and the GPU heatsink, which can be tricky without a 3D printer or proper templates. If you want to minimize risk to the card and the blower bearings, your safest bet—though more expensive—is buying the official V620 blower from AliExpress, since it’s designed to mount correctly and keep the Precision spindle bearings operating within spec.
This site uses cookies to help personalise content, tailor your experience and to keep you logged in if you register.
By continuing to use this site, you are consenting to our use of cookies.
