Bought a similar months ago for 150USD shipped from US. Broken after 1week running passed through to 7 Win vm's. Now what? I prefer to buy from eBay no matter of the sellers reputation over buying here from 'good seller' and STH member from Hungary and being ripped off. With eBay I always get my money back.
Tesla P4 - 8GB / US$ 100 or even lower
- Thread starter gb00s
- Start date
Notice: Page may contain affiliate links for which we may earn a small commission through services like Amazon Affiliates or Skimlinks.
I use on of these for video transcoding and AI inference on security camera footage. It pulls less than 25W on that workload, and for Plex transcoding, for example, is a much better value than the frequently suggested P2000/P2200. It's essentially a frequency nerfed 1080, but with 8GB of ECC memory, and in a smaller package. Supports vGPU out of the box, and with vgpu_unlock-rs you can work around some of the limitations. Decent gaming performance too, though if that's your primary use case you can probably find a equal or better performing GeForce card at the same price.
Not sure why some folks here say it's useless. It's kind of a hidden gem.
Not sure why some folks here say it's useless. It's kind of a hidden gem.
This is a wild comment. The Tesla P4 is a 75w TDP, half-height card w/ 8GB memory. The K80 is a slower, more expensive, 300w TDP, full-height / 2-slot card w/ 24GB memory.
Forget the K80 part then. Just saying if you can fit a full height card in there are other alternatives worth looking at that are better value...
Please name such a card. I’m on the market for a low power VM gpu and the Tesla P4 was the card I was probably going to get. If you’ve got something that renders the P4 e-waste I’d love to know what it is.
I used the P4 for exactly this purpose too - video transcoding and have found it to be very cost effective. US $100 is a steal in my opinion if its used for this purpose
OK. I was under the initial impression people were getting this for games as the craft computing video and eduncan mentioned. For those 2U servers with no extra power port (or for those of us with a limited power budget in our colocation...), this thing is absolutely perfect for NVENC and hardware acceleration for a virtual desktop, and I own this card for that purpose. Just don't expect miracles from it in anything else...
There aren't many benchmarks out there for the P4 (and craft computing conveniently didn't compare this to ANY consumer card) but we can see the T4 (which is about 7 times the cost) performs similar to a 1070 or 2060 in both gaming and what it was designed to do - inference. https://www.servethehome.com/nvidia-tesla-t4-ai-inferencing-gpu-review - so I would expect in gaming the P4 to perform similar or worse than a 1060 or 9xx series card for those tasks. The huge benefit is the efficiency at which it does so. These cards are designed for power efficiency. The extra video memory is very useful as well for virtual desktops.
Another thing is it's a bitch to cool, you need good airflow, or one of those 3D printed shrouds.
There is also the GTX 1650 low profile if you want a slightly more modern (Volta), similarly priced, desktop GPU that is also 75W/low profile and not a pain to cool. I imagine it will perform a little bit better, but at the disadvantage of 4GB less VRAM.
But again if you have a case that allows a full height GPU and you do care about games performance, don't get this, you will be disappointed. Once you also buy the shroud and fan to cool it you might as well have purchased a 1070.
Also since I'm sick of not being able to find benchmarks for this card, if I have time today I will run the 3Dmark suite on a P4 and post the results. I would expect similar to 1650 performance.
Last edited:
Anything working technically isn't ewaste--it's just needing to be matched with a use-case. E-waste is stuff that is broken that needs to be refurbed or broken back down for further recycling.
Saying older working stuff is e-waste is like saying someone who is ignorant is stupid--they're not the same thing.
Awfully roundabout way of saying I'm an idiot for posting that
I clarified above, it's not great for gaming, other use cases, go for it.
Did you not use something like paypal that has some sort of protection?
Very few people here would be using a card like this for games as most do not use their lab for gaming. The audience is always something to think about when discussing video cards.
I'm using this:
github.com
with Looking-Glass remotely via RocE (RDMA) with this:
github.com
So I have virtualized my CAD Workstation - and yes I tried gaming - working really good so far
So no need for old "Enterprise" Cards anymore (except if you want power efficenzy with big VRAM (but u could underclock & undervolt "old" Titan cards as well))
GitHub - mbilker/vgpu_unlock-rs: Unlock vGPU functionality for consumer grade GPUs
Unlock vGPU functionality for consumer grade GPUs - mbilker/vgpu_unlock-rs
github.com
GitHub - telescope-proj/lgproxy: Proxy for Looking Glass over local networks
Proxy for Looking Glass over local networks. Contribute to telescope-proj/lgproxy development by creating an account on GitHub.
github.com
So I have virtualized my CAD Workstation - and yes I tried gaming - working really good so far
So no need for old "Enterprise" Cards anymore (except if you want power efficenzy with big VRAM (but u could underclock & undervolt "old" Titan cards as well))
Last edited:
Seller contacted me today.
Apologized for the listing indicating that the cards have Low Profile brackets. While the cards (and pictures on ebay) have High Profile brackets.
Offered to send me the low profile brackets in a separate shipment.
So they are communicating and looking to make good on what they offer.
According to the Track and Trace I should receive my cards on Friday. LP brackets are being sent by snail mail so will take a bit longer.
Will update after I have received the cards.
Apologized for the listing indicating that the cards have Low Profile brackets. While the cards (and pictures on ebay) have High Profile brackets.
Offered to send me the low profile brackets in a separate shipment.
So they are communicating and looking to make good on what they offer.
According to the Track and Trace I should receive my cards on Friday. LP brackets are being sent by snail mail so will take a bit longer.
Will update after I have received the cards.
Do you need to purchase the Nvidia vGPU licenses when using consumer cards?I'm using this:
with Looking-Glass remotely via RocE (RDMA) with this:GitHub - mbilker/vgpu_unlock-rs: Unlock vGPU functionality for consumer grade GPUs
Unlock vGPU functionality for consumer grade GPUs - mbilker/vgpu_unlock-rs
GitHub - telescope-proj/lgproxy: Proxy for Looking Glass over local networks
Proxy for Looking Glass over local networks. Contribute to telescope-proj/lgproxy development by creating an account on GitHub.
So I have virtualized my CAD Workstation - and yes I tried gaming - working really good so far
So no need for old "Enterprise" Cards anymore (except if you want power power efficenzy with big VRAM (but u could underclock & undervolt "old" Titan cards as well))
lol! That wasn't my intention. It's just that so many people in our hobby think 'old' = 'ewaste' and spread this like some sort of bible to follow. And unfortunately there's a lot of people that follow this terrible advice, buy something more 'appropriate' and realize it's a massive overkill and waaay too expensive of a hobby when if they would have started small, they would have 'stepped up' via planned and budgeted upgrades that fit a need.Awfully roundabout way of saying I'm an idiot for posting that
I clarified above, it's not great for gaming, other use cases, go for it.
I'm not saying that old can't be ewaste--just imagine all the cell phones that are 2g or earlier that were perfectly working, but can't work because there's no longer these signals anywhere on the planet. I'm sure the cream of of them will be good for museum pieces but the rest will need to be recycled. I actually have a uniden bag cell phone from this era somewhere--it was so high quality too. We never used it, lol. What was awesome about it though is that it shared the same battery as our full size VHS camcorder, so both had spare batteries. But even VHS is nearly gone, even for archival use.
I am interested to know what non-gaming use-cases people have for Tesla cards like these in homelabs. I am thinking about getting one to play around with in a Supermicro Haswell-era system with no GPU / iGPU and Proxmox. Full disclosure, I am kinda looking for a reason to get one... but I am a novice regarding vGPUs and VDI .
If I understand correctly, vGPUs can be configured on cards like this P4 and assigned to VMs, much like how vfio-passthrough would work with a physical GPU, only the Teslas have no physical graphics output. So, applications on the VM like Plex or Blue Iris could leverage hardware acceleration for video encoding / decoding, as long as the VM OS has drivers for the card and the application is aware of it.
In contrast, to use a vGPU in a desktop environment, one would need to configure a VDI software like moonlight or Parsec, but not regular mstsc.com or VNC, and connect from another client. Is that correct? In this scenario the vGPU encodes using h.264 / h.265 and the client only needs to decode that codec to have near-native capability? Can a vGPU be assigned to a LXC container or just to VMs?
.If I understand correctly, vGPUs can be configured on cards like this P4 and assigned to VMs, much like how vfio-passthrough would work with a physical GPU, only the Teslas have no physical graphics output. So, applications on the VM like Plex or Blue Iris could leverage hardware acceleration for video encoding / decoding, as long as the VM OS has drivers for the card and the application is aware of it.
In contrast, to use a vGPU in a desktop environment, one would need to configure a VDI software like moonlight or Parsec, but not regular mstsc.com or VNC, and connect from another client. Is that correct? In this scenario the vGPU encodes using h.264 / h.265 and the client only needs to decode that codec to have near-native capability? Can a vGPU be assigned to a LXC container or just to VMs?
You could spoof the PCIe IDs of the mdev, to match the IDs of a Quadro, so no need for grid driver, normal Quadro Driver works for Windows.
Otherwise you could do a each day grid driver license reset for the grid driver, to use the mdev with "official" grid driver.
vgpu_unlock does work with series 90, 10 and 20 - for now, not working with 30 or 40
(except for cards with two memory regions like gtx970)
A mdev could probably be used in LXC or docker/podman - but I have not tried yet
Only when placed in a desktop or workstation. Mine goes into the server ofc.
There is no Quadro P4. However, this hugly depends on your software. You probably wont see any benefit if your output is different. For rendering or video encoding you can choose a dedicated gpu in most cases.
Last edited: