I wonder if you could use a larger GPU that would otherwise need a full-size case and mount the Pi inside the GPU? More realistic probably is just 3d printing a test bench wrapper for your PSU; mount the Pi to the side and attach the graphics card to the top somehow on its side or something.
Many ARM devices are small enough to be build into the eGPU stands base. Combine that with a very compact 350W range SFX or TFX power supply, and you'd be golden.
@@Genesis8934 Yeah. This can be easily done too with any 3D printing slicer software. You can just grab the individual .stl models for each item, and then in the slicer join all models together. It may not look the best doing it this way, and you could run into weird issues with support, but as long as everything is a simple shape, its fairly simple. Ive done that for almost this same reason. I have a Minisforum MS-01 that I needed to run a PCIe-x8 card in that required power, so I found a .stl file that was an external case for an ATX PSU that had basically an endcap that kept all the cabled bundled up, while only allowing the ones you needed to pass through. I then joined a .stl for a PCIe-x16 riser / vertical mount GPU ribbon to that PSU print, so now the card is mounted and stable attached to the PSU case while the ribbon runs into the back of the MS-01.
This setup is fascinating! It’s incredible to see the evolution of Pi hardware to the point where external GPUs are not only viable but practical. Your deep dive into the AMD GPU patch and Polaris cards shows just how far open-source drivers and Pi’s PCIe bus have come. The mix of gaming benchmarks with Doom and 4K transcoding possibilities makes this more than just a hobby project-it’s a legit, low-power home lab solution. I’d be curious to see how this rig would handle Blender or OBS for media creation and whether future patches can further optimize that PCI bandwidth bottleneck.
I've so far found a lot of applications assume that arm64 == no possibility of having a GPU, so some things will have to wait for software devs to wake up to the fact that arm64 is a serious computing platform these days!
When I first watched this video I didn't understand what and how are those connectors/adapted configured for this to work, but then I went to your blog post and read it, thanks for documenting this in that way! Super B video
I@@JeffGeerlingI'm a noobie to all of this, and a very old man too boot I'd love to read your blogs Jeff but I don't know how to find where they reside, brilliant video by the way 😊
Sweet Mother of Cable! I remember watching your older videos with the previous generation pi model, trying to figure out if there was a way to utilize an external graphics card. So happy it is finally possible with the new model and the help of others. I bet it was a quite an exciting and moving moment when the setup showed its first functioning picture/screen.
Hey Jeff! Congrats on the baby and glad to see you back on YT. Really like your channel. I don’t always understand what you’re talking about but I still enjoy your content! Appreciate it. 👍
I've been following along your progress with compiling gpu drivers for the pi and i gotta say, it's impressive that it finally works. mad respect to the FOSS community
Doom 3 has such a goated engine. I spent some time developing for idTech 4 and it’s amazingly modern for a 24 year old engine. Also it’s open source! So you can custom fit it to basically any architechture . Great video btw!
There are plenty of usecases for GPU with RPi5. Emulation will benefit greatly from the added more powerful GPU, you can make steam machine with some additional translation layers but for steam link and some indie or older titles this should be plenty, mini 3D workstation sounds good too especially if you are into 3D printing since new blender versions take advantage of GPU power lot more than older versions. Photo and video editing software will benefit from more GPU power as well. The drivers and installation process needs some work but I'm sure it will get there. This is great proof of concept though.
This push came from you and us in the comments and the videos and projects on/off Github made it all possible. Thank you Jeff for this progress and this demonstration of elite edge computing.
I just had a look at Doom III and it freaked me out! 'Welcome to Mars' but down in the dungeons everyone looks like they've got lupus. Fantastic ray-tracing, even today it looks like something out of the future. 🖋
6:02 Actually, according to an internal steam change that steamdb managed to pick up, valves been playing around with arm, along with tools like fex and waydroid
Just a mini tip, check if your pi was not the bottleneck for GPU by using something like BTOP. It monitors *everything*, and I think (assumption) that your GPU is being bottlenecked by your CPU in Doom 3, hence it was not being 100% utilized. notation: This is not 100% coherent, just an assumption. If Jeff sees this, I'd be curious to know if it is being 100%ed, tho I don't remember if doom 3 is multithreaded. Thanks in forward! :)
The CPU wasn't the bottleneck in Doom 3's case - it was only maybe 30-50% loaded. btop is awesome for that though! Doom Eternal was getting CPU bottlenecked, though. There may be some driver optimizations to remove that bottleneck, as a lot of it is in memory address translation.
Btop is great, I love having it open on a second monitor while doing something cpu/gpu/io intensive and instantly being able to assess the whole system load. I always wonder what other programs like htop I have used for decades not realising there is a well supported superior alternative. I felt the same way upon discovering tmux ~2010 having always used screen.
@@beardedchimp Yes, btop is awesome! Also s-tui instead of just using stress-ng and having another monitor monitoring things. Lots of cool tools that I find out about through these comments!
Next step is a case that will take a Pi5, ATX power supply and a pcie slot adapter. Better yet, once CM5 is released, you could bake a bunch of that into a board and slap the compute module on.
i don't have the recompiled the linux kernel shirt but i have the mug instead, really cool to see this project going somewhere, would be fun to toy with a pi myself
Hey Jeff, I've heard using the pi's composite out slows the pi down because of frame rate or something. How would it work if you used this setup with an older card with DVI-I connected to a CRT? I'm wondering if that would slow the pi down? I'm guessing it wouldn't.
I love this. But I mainly use Macs and don't understand PCIE or M.2 or all the other adaptors. The last time I build a PC was 1998. I kind of want someone to make a large Pi case that just has a slot for a GPU and a single power supply that does everything so I can just plug & play and not have to have a janky setup with all kinds of adaptors.
this is cool as fawk and opens up alot of doors to making a very cool mini gaming pc. I have a old rx 580 laying about with some of those pcie mining slots that connect to SATA
@@autohmae i mean these things already existed in the past before hardware accelerated video blocks where on CPUs and GPUs, im just asking for them to make a comeback with modern standards :P
Have u ran any tests with dedicated 4k capable games? I know the original doom 3 was made so long ago it wouldn’t be a dedicated 4k game but obviously upscale. Maybe games like cyber punk, zen pinball actual rendered 4k games and see how they perform
Depends mostly on what format of videos you have, plus what transcoding settings you set up - I ran into issues a lot if I had 4K UHD BR rips, and tried streaming them to my phone at 1080p.
I wonder how good arm and a Nvidia GPU combo could be maybe in a couple years we will be able to play games under windows that way and have better performance if possible?
8:17 I think it does. I think you can either disable this in settings or it might be in the chrome flags page. That said, I don't know if the Pi CPU is capable of running Chrome without hardware accel lol (maybe!).
Hi Jeff! Thanks for all the great videos. A tangential question; what do you do to protect your computers from EM interference? Recently solar storms have caused some strong surface EM interference. This has caused some down time for many, even big companies. I'm just wondering if some aluminium foil would reduce these effects or if a stronger solution is needed. Thanks again!
Big big favor to ask. Please do a video with the pi 5 security camera using Shinobi an the latest highj res camera and hostapd with a usb wifi dongle as part of a mesh network. Dual purpose goodness with a raspberry pi
Very cool! It's nice to see those games running on alternative hardware. Also, as for Jellyfin, I run it on an ODROID H4 which has an Intel chip. According to Jellyfin Intel has the best transcode quality and the N97 can do like 2 4K transcode streams in realtime without any issues.
I'm curious. Have you tested Wine with it and can anything run under that? This also reminds me of when I was testing out my Snapdragon X Elite with an eGPU attached to it. Sadly no Nvidia drivers for Windows on ARM to make it happy but it at least did detect it.
I've been testing many LSI cards - it seems like the 93xx line is the first one that supports arm64 architecture in the drivers, so you have to get one of those I think, or newer.
@@JeffGeerling That makes sense. No point creating drivers for products which are ancient in tech terms. At least prices have dropped massively for used 93xx cards since I last looked. I'll keep a look out for your future Lsi video. Actually, I watch all your videos, I've become a Geerling Geek .
Chrome is trying to use VAAPI as your HWaccel, hence the issues. Are these cards new enough to support vulkan or OpenCL as your HWaccel? Arch has a package in the AUR called opencl-amd that's just the opencl part of ROCM without having to install ROCM. The rpi5 does not have video encoding hardware built into the CPU, so it is not practical to be used an encode server. But a low power/cost arm encode server would be a great use for some people.
The lower power costs are more then offset by the price of a pi5 and the adapter / GPU contraption. For over $150 I could get a lot of power. Or a even faster and still efficient AMD Ryzen mini PC / intel nuc.
Getting an old Dell/HP/LEnovo with a 7th Gen Intel and using quicksync for the video transcoding is probably the better option. But good to know some of the options out there.
what I imagine is a very cheap SBC that features m.2 + a cheap mining adapter (m.2 -> usb -> 1x slot) + GPU + cheap switching 12v PSU (400w or so would power GPU and SBC) and that will reduce the whole rig price so much. Now 3d print some basic mounts and you have a single gpu mining rig for like 60~ bucks. Scaling from there is just replicating the setup, you can use a lot of deployment features that will get you from a brand new sbc to a fully running and configured rig very quickly. If anything fails it's always a very cheap replacement, and the whole thing would be very easy to monitor. If only gpu mining was worth it on late 2024 lol.
Raspberry PI question. Can I connect a Raspberry Pi 5 Penta SATA HAT and a PCIe to M.2 Adapter with PoE Function with a HatNET! 2.5G (2.5 Gigabit Ethernet) and a HatBRICK! Commander (2 Ports, GEN 2) to a Raspberry PI 5?
This will only get easier to play games on things like the raspberry pi with Proton ARM. Something Valves working on but i don't think much is public yet.
Are you not able to use a more recent version of mesa? in the screenshots and on the stream it seemed like you were using 23.2 which is like a year old and there have been a lot of improvements since, both for amd gpus and the rpi 5 gpu. Also you should be able to use OpenCL via rusticl with the rx 460 if you set a environment variable (though it would probably work better if you are able to update to an newer version of mesa as it had just been added in 23.1)
When I was growing up, I was an avid gamer, and me and my friend both agreed, that the playability of a game is paramount, and graphics basically don't matter. These types of computers, are for establishing a grid, where there isn't one at all. If you really wanted better graphics, you could go outside. I work with airplanes, and I would love a 200$ computer that I could use to run training programs on various airplanes. You don't need good graphics for that, you just need accurate representation of physics. If I wanted more abilities, I'd add more motherboards, and set each to run a seperate program, before vectoring all the data, into a filter, to compile it, and display it.
Next up has to be something like Over:board, a CM4 Mini-ITX motherboard. Then we can hope the RPI foundation will use the same form factor for the CM-board in the future!
do a video on the most cost effective (possibly pi based?) homelab media server. should have 4 drive NAS with raid and also video transcoding so that it supports all modern videos (my pi5 works for jellyfin but some video formats are jittery) so then I could potentially have an all-in-one nas with a gpu, serving video and doing some ai tasks! otherwise I am thinking of just building a computer in a case, but it would be way more expensive... looking at most cost effective solutions and I already have a pi5 doing most of the job!
I heard you could force Gen 3 pcie speeds. The pi 5 is certified for gen 2 speeds, but is there a reason for not allowing it to run at Gen 3 speeds? Is it a stability issue, or would it not be guaranteed to hit the full gen 3 speeds ?
I've actually had no issue getting LLM's to run on older graphics cards and with Llama 3.2's models being under 2GB you can spin up a baremetal LLM with Ollama without too much issue (assuming the software support is there). I would be very curious if it would be possible to get a slightly more modern AMD GPU that's Ollama supported running, especially once multi-modal support is in it could make for an interestingly sized computer vision system.
When I check the release page, they have ollama-linux-arm64. Then again, an LLM can also still be CPU bound even if they have enough GPU, so I think that might be a problem.
Instead of mounting the GPU to the Pi, mount the Pi to the GPU. Absolute structural stability.
4D Pi mounting options!
@@JeffGeerling just mount the pi in the future so you don't have to worry about it now!
I wonder if you could use a larger GPU that would otherwise need a full-size case and mount the Pi inside the GPU?
More realistic probably is just 3d printing a test bench wrapper for your PSU; mount the Pi to the side and attach the graphics card to the top somehow on its side or something.
Many ARM devices are small enough to be build into the eGPU stands base. Combine that with a very compact 350W range SFX or TFX power supply, and you'd be golden.
@@Genesis8934 Yeah. This can be easily done too with any 3D printing slicer software. You can just grab the individual .stl models for each item, and then in the slicer join all models together. It may not look the best doing it this way, and you could run into weird issues with support, but as long as everything is a simple shape, its fairly simple.
Ive done that for almost this same reason. I have a Minisforum MS-01 that I needed to run a PCIe-x8 card in that required power, so I found a .stl file that was an external case for an ATX PSU that had basically an endcap that kept all the cabled bundled up, while only allowing the ones you needed to pass through. I then joined a .stl for a PCIe-x16 riser / vertical mount GPU ribbon to that PSU print, so now the card is mounted and stable attached to the PSU case while the ribbon runs into the back of the MS-01.
Hard work should be rewarded. Here you go. You inspire me.
W
W
Now it's time to fire up the Pi to watch Jeff Geerling in 4K play games in 4K on a Pi.
Or do 8K upscaling with FSR! Haha
Pi ception
Writing this comment on a 1B
Totaly Meta bruh...
Totally Meta bruh...😂
Swedish computer site Sweclockers was reporting on your last Pi GPU video 2 days ago. You are a legend Jeff.
Your studio, you need a “recompile kernel counter” in the background.
Or a counter like the ones at logistics companies and factories.
"0 days without recompile"
This setup is fascinating! It’s incredible to see the evolution of Pi hardware to the point where external GPUs are not only viable but practical. Your deep dive into the AMD GPU patch and Polaris cards shows just how far open-source drivers and Pi’s PCIe bus have come. The mix of gaming benchmarks with Doom and 4K transcoding possibilities makes this more than just a hobby project-it’s a legit, low-power home lab solution. I’d be curious to see how this rig would handle Blender or OBS for media creation and whether future patches can further optimize that PCI bandwidth bottleneck.
I've so far found a lot of applications assume that arm64 == no possibility of having a GPU, so some things will have to wait for software devs to wake up to the fact that arm64 is a serious computing platform these days!
Just wait for the USB4 on smartphones. Docking them will unlock a lot of options for us.
@@JeffGeerlingblame Microsoft on this
When I first watched this video I didn't understand what and how are those connectors/adapted configured for this to work, but then I went to your blog post and read it, thanks for documenting this in that way!
Super B video
Blog > video
(Except only like 1/1000th of my audience here reads blogs!)
I@@JeffGeerlingI'm a noobie to all of this, and a very old man too boot I'd love to read your blogs Jeff but I don't know how to find where they reside, brilliant video by the way 😊
@@johnbowles4754there's a link in the description. Click "…more" in the area under the video and you'll see it. 👍
@@rogerramjet8395 thank you so much 💗
@@johnbowles4754 I'll often link to a relevant article in the description-but it's all at www.jeffgeerling.com :)
Sweet Mother of Cable! I remember watching your older videos with the previous generation pi model, trying to figure out if there was a way to utilize an external graphics card. So happy it is finally possible with the new model and the help of others. I bet it was a quite an exciting and moving moment when the setup showed its first functioning picture/screen.
Hey Jeff! Congrats on the baby and glad to see you back on YT. Really like your channel. I don’t always understand what you’re talking about but I still enjoy your content! Appreciate it. 👍
I've been following along your progress with compiling gpu drivers for the pi and i gotta say, it's impressive that it finally works. mad respect to the FOSS community
I didnt think I'd see the day that these little SBCs would be able to use a external GPU - this is epic.
Thanks Jeff. I am thoroughly impressed you got that video card mounted on the raspberry Pi 5!
I am learning a lot from your blogs and videos. Thanks.
Someone is gonna make a portable Mining RIG and hide it in some public place, stealing electricity.
Solar might be better for that.....
Is gpu mining still a thing? Bitcoin only has value for online casinos.
🤫
Aka a gaming laptop.
@@LoganE01 It will cost a lot, You will need a decent battery atleast.
Doom 3 has such a goated engine. I spent some time developing for idTech 4 and it’s amazingly modern for a 24 year old engine. Also it’s open source! So you can custom fit it to basically any architechture .
Great video btw!
Using “goated” really showcases your mid 60s IQ
@@stcojo awww, who hurt you little man? Mommy yelled at you again? :(
@@TuRmIx96 wipe your tears, i won
@@stcojo 👏👏👏👏👏👏 congratz, mom is proud ;)
It took a really long time but you finally did it !!!
This is an amazing video, thank you so much for your Input!
so much effort finally offered results, impressive work
i didnt knew lian li made psus
Jeff you are paving the way for the future of smartphones. Once we have USB4 on them docking with eGPU would be the dream for the living room setup.
This is awesome. This kind of work will always get a like from me. 😁
There are plenty of usecases for GPU with RPi5. Emulation will benefit greatly from the added more powerful GPU, you can make steam machine with some additional translation layers but for steam link and some indie or older titles this should be plenty, mini 3D workstation sounds good too especially if you are into 3D printing since new blender versions take advantage of GPU power lot more than older versions. Photo and video editing software will benefit from more GPU power as well.
The drivers and installation process needs some work but I'm sure it will get there. This is great proof of concept though.
This push came from you and us in the comments and the videos and projects on/off Github made it all possible. Thank you Jeff for this progress and this demonstration of elite edge computing.
I just had a look at Doom III and it freaked me out! 'Welcome to Mars' but down in the dungeons everyone looks like they've got lupus. Fantastic ray-tracing, even today it looks like something out of the future. 🖋
Jeff is always coming up with the funnest projects
It would be cool to have a variant of the Pi that has a dedicated graphics chip onbord!
6:02 Actually, according to an internal steam change that steamdb managed to pick up, valves been playing around with arm, along with tools like fex and waydroid
No power supply paper clip trick 😂?
Heh after accidentally shorting out a 12v rail one time I decided to buy the nice hardware switch, with insulation!
@@JeffGeerling that’s entirely too reasonable sir 😉
@@JeffGeerling if you do these things often... that sounds like the superior way to do it
Once did it with a random piece of wire I found in a drawer!
I'd love to see your videos working on the drivers! Low level programming babyyyy!
Bro finally did it
Just a mini tip, check if your pi was not the bottleneck for GPU by using something like BTOP. It monitors *everything*, and I think (assumption) that your GPU is being bottlenecked by your CPU in Doom 3, hence it was not being 100% utilized.
notation: This is not 100% coherent, just an assumption. If Jeff sees this, I'd be curious to know if it is being 100%ed, tho I don't remember if doom 3 is multithreaded. Thanks in forward! :)
The CPU wasn't the bottleneck in Doom 3's case - it was only maybe 30-50% loaded. btop is awesome for that though!
Doom Eternal was getting CPU bottlenecked, though. There may be some driver optimizations to remove that bottleneck, as a lot of it is in memory address translation.
Btop is great, I love having it open on a second monitor while doing something cpu/gpu/io intensive and instantly being able to assess the whole system load. I always wonder what other programs like htop I have used for decades not realising there is a well supported superior alternative. I felt the same way upon discovering tmux ~2010 having always used screen.
@@beardedchimp Yes, btop is awesome! Also s-tui instead of just using stress-ng and having another monitor monitoring things. Lots of cool tools that I find out about through these comments!
Good to see you finally succeeded
Next step is a case that will take a Pi5, ATX power supply and a pcie slot adapter. Better yet, once CM5 is released, you could bake a bunch of that into a board and slap the compute module on.
This is... amazing!! 🤩
i don't have the recompiled the linux kernel shirt but i have the mug instead, really cool to see this project going somewhere, would be fun to toy with a pi myself
With some PCI-E magic to run an HBA along with the GPU, you could have a pretty snappy Jellyfin instance running completely on a Pi. Pretty wild.
That's super cool!) RPi5 possibility is growing)
Hey Jeff, I've heard using the pi's composite out slows the pi down because of frame rate or something. How would it work if you used this setup with an older card with DVI-I connected to a CRT? I'm wondering if that would slow the pi down? I'm guessing it wouldn't.
we got 4k gameplay on raspberry before gta 6
haha
Jeff waiting for the year of egpu on pi feels more passionate than the endless year of linux desktop 😹
2:35 Or you could just short the green wire to GND. They're all black in modern PSUs but you can find the pin diagram on the internet.
Well done Jeff, well done! 👏👏👏
7:14 I can report I've had trouble with hardware 10-bit encoding on the RX 570 in my pc as well, not sure what the issue is
Huh, maybe driver support then! Sometimes hardware specs don't match up to what actually works :(
Let's go! Also great video cheers
Seems like a lot of hardware for a resolution I never really miss that much. I wish there was a better hat.
I love this. But I mainly use Macs and don't understand PCIE or M.2 or all the other adaptors. The last time I build a PC was 1998. I kind of want someone to make a large Pi case that just has a slot for a GPU and a single power supply that does everything so I can just plug & play and not have to have a janky setup with all kinds of adaptors.
Ngl, wiuld be interesting to seearm sbc solutions in the future, that would support gpus out of the box
this is cool as fawk and opens up alot of doors to making a very cool mini gaming pc. I have a old rx 580 laying about with some of those pcie mining slots that connect to SATA
ngl ,i kinda just want an AV1 hardware encoder USB stick that i can plug into my Pi for jellyfin
USB-sticks exists for AI exist, so maybe that could be a thing in the future
@@autohmae i mean these things already existed in the past before hardware accelerated video blocks where on CPUs and GPUs, im just asking for them to make a comeback with modern standards :P
@@Mobile_Dom OK, well, I was just pointing out, some other things are already done this way. So it's not much of a stretch both ways
@@autohmae oh agreed, i didnt mean to some off as combative
@@Mobile_Dom no, don't worry, just that we understand each other 🙂
Love the T-shirt 😂 I’m a security guy so I get it. 👍
Could we get a kernel with Intel ARC support? AV1 transcoding on the cheap would be really cool.
Intel has some open source drivers I haven't taken for a spin yet. Though they seem to have petered out releasing new GPUs lately :(
Have u ran any tests with dedicated 4k capable games? I know the original doom 3 was made so long ago it wouldn’t be a dedicated 4k game but obviously upscale. Maybe games like cyber punk, zen pinball actual rendered 4k games and see how they perform
I’ve never had any issues with buffering on Jellyfin using standard Raspberry Pi and an external 4TB spinning rust hard drive with a USB 3.0
Depends mostly on what format of videos you have, plus what transcoding settings you set up - I ran into issues a lot if I had 4K UHD BR rips, and tried streaming them to my phone at 1080p.
This is actually really cool
oh wow and isnt proton working on arm
Wonder how if at all possible that would work on something like a Radxa Rock 5 ITX
I wonder how good arm and a Nvidia GPU combo could be maybe in a couple years we will be able to play games under windows that way and have better performance if possible?
Seems like the next step would be trying to get Bazzite with steam proton working on it.
It finally happened!
Super exciting $300 octopus pi setups for jellyfin transcoding 😂😂😂
My newest gpu is probably 11 years old, I do want to try this out when I get a chance though
8:17 I think it does. I think you can either disable this in settings or it might be in the chrome flags page. That said, I don't know if the Pi CPU is capable of running Chrome without hardware accel lol (maybe!).
Hi Jeff! Thanks for all the great videos. A tangential question; what do you do to protect your computers from EM interference? Recently solar storms have caused some strong surface EM interference. This has caused some down time for many, even big companies. I'm just wondering if some aluminium foil would reduce these effects or if a stronger solution is needed. Thanks again!
Big big favor to ask. Please do a video with the pi 5 security camera using Shinobi an the latest highj res camera and hostapd with a usb wifi dongle as part of a mesh network. Dual purpose goodness with a raspberry pi
Very cool! It's nice to see those games running on alternative hardware.
Also, as for Jellyfin, I run it on an ODROID H4 which has an Intel chip. According to Jellyfin Intel has the best transcode quality and the N97 can do like 2 4K transcode streams in realtime without any issues.
I'm curious. Have you tested Wine with it and can anything run under that? This also reminds me of when I was testing out my Snapdragon X Elite with an eGPU attached to it. Sadly no Nvidia drivers for Windows on ARM to make it happy but it at least did detect it.
Hats off to Coreforge!
Do you think you'll have time to setup popular old reliable LSI storage cards working on the Raspberry Pi 5? Like the LSI 9211-8i.
I've been testing many LSI cards - it seems like the 93xx line is the first one that supports arm64 architecture in the drivers, so you have to get one of those I think, or newer.
@@JeffGeerling That makes sense. No point creating drivers for products which are ancient in tech terms. At least prices have dropped massively for used 93xx cards since I last looked. I'll keep a look out for your future Lsi video. Actually, I watch all your videos, I've become a Geerling Geek .
Chrome is trying to use VAAPI as your HWaccel, hence the issues. Are these cards new enough to support vulkan or OpenCL as your HWaccel? Arch has a package in the AUR called opencl-amd that's just the opencl part of ROCM without having to install ROCM. The rpi5 does not have video encoding hardware built into the CPU, so it is not practical to be used an encode server. But a low power/cost arm encode server would be a great use for some people.
omg you finally did it!
The lower power costs are more then offset by the price of a pi5 and the adapter / GPU contraption. For over $150 I could get a lot of power. Or a even faster and still efficient AMD Ryzen mini PC / intel nuc.
Getting an old Dell/HP/LEnovo with a 7th Gen Intel and using quicksync for the video transcoding is probably the better option. But good to know some of the options out there.
I've been really curious to how well games work with box86 on the Pi with a GPU.
I would have thrown an RX580 8GB in the slot, they are cheap and packs quite a decent little punch.
You should cut an oculink cable in half and solder it to custom FPC cable!
Maybe even have stiffener on the FPC with a hole to screw the FPC to the PI standoff.
Heh, definitely room for improvement in terms of the 'jank factor' here!
I'm thinking an Oculink HAT might be useful at least.
Can you test few parallel GPU cards whether they can run llama AI models using rpi?
Thanks for that video! Do you know if Nvidia GPU also work on raspberry pi 5?
what I imagine is a very cheap SBC that features m.2 + a cheap mining adapter (m.2 -> usb -> 1x slot) + GPU + cheap switching 12v PSU (400w or so would power GPU and SBC) and that will reduce the whole rig price so much. Now 3d print some basic mounts and you have a single gpu mining rig for like 60~ bucks. Scaling from there is just replicating the setup, you can use a lot of deployment features that will get you from a brand new sbc to a fully running and configured rig very quickly. If anything fails it's always a very cheap replacement, and the whole thing would be very easy to monitor.
If only gpu mining was worth it on late 2024 lol.
Cannot wait to play supertuxkart on a 50 dollar sbc on my rx 6942 that costs 800 dolla
what's the name of your gpu monitoring cli tool?
found the answer at 5:34 :p
@@sonickautschuhe nvtop, that's it!
Your info is priceless.
The next trick is to get an RTX4090 working on a C64.
Within the Mesa project there's Rusticl which provides OpenCL 3 for compute on modern AMD GPUs, OpenCL support for Mesa is usually packaged separately
And here i was, all proud of myself for finally getting my Lubuntu laptop's sdcard reader working again
Would love to help! Ive got an Amd GPU and some working knowledge in Kernel drivers
Raspberry PI question. Can I connect a Raspberry Pi 5 Penta SATA HAT and a PCIe to M.2 Adapter with PoE Function with a HatNET! 2.5G (2.5 Gigabit Ethernet) and a HatBRICK! Commander (2 Ports, GEN 2) to a Raspberry PI 5?
Great work Jeff
Make a 3D housing for it to house both the GPU and the Pi on top!
This will only get easier to play games on things like the raspberry pi with Proton ARM. Something Valves working on but i don't think much is public yet.
Are you not able to use a more recent version of mesa? in the screenshots and on the stream it seemed like you were using 23.2 which is like a year old and there have been a lot of improvements since, both for amd gpus and the rpi 5 gpu.
Also you should be able to use OpenCL via rusticl with the rx 460 if you set a environment variable (though it would probably work better if you are able to update to an newer version of mesa as it had just been added in 23.1)
If SteamOS releases, this is going to spawn a whole market of small game stations.
Such a cool idea.
Maybe we could compare the Raspberry Pi and Orange Pi.
Orange Pi 5 Max is pretty awesome little hardware. Support side is a bit different, but if you can make use of it, it's a better value.
When I was growing up, I was an avid gamer, and me and my friend both agreed, that the playability of a game is paramount, and graphics basically don't matter.
These types of computers, are for establishing a grid, where there isn't one at all.
If you really wanted better graphics, you could go outside.
I work with airplanes, and I would love a 200$ computer that I could use to run training programs on various airplanes.
You don't need good graphics for that, you just need accurate representation of physics.
If I wanted more abilities, I'd add more motherboards, and set each to run a seperate program, before vectoring all the data, into a filter, to compile it, and display it.
Cool. The sound in this video sounds a bit weird, it's like a background noise.
Next up has to be something like Over:board, a CM4 Mini-ITX motherboard. Then we can hope the RPI foundation will use the same form factor for the CM-board in the future!
do a video on the most cost effective (possibly pi based?) homelab media server. should have 4 drive NAS with raid and also video transcoding so that it supports all modern videos (my pi5 works for jellyfin but some video formats are jittery) so then I could potentially have an all-in-one nas with a gpu, serving video and doing some ai tasks!
otherwise I am thinking of just building a computer in a case, but it would be way more expensive... looking at most cost effective solutions and I already have a pi5 doing most of the job!
that chrome issue happens on my ryzen cpu with amdgpu. So i dont think its arm related.
Ha! Glad it's not just me then.
2:20 did you wire up the PSU
Just ask can it run a xtx of 24gigs of vram... could be useful for ai stuff... and games/emu's
I heard you could force Gen 3 pcie speeds. The pi 5 is certified for gen 2 speeds, but is there a reason for not allowing it to run at Gen 3 speeds? Is it a stability issue, or would it not be guaranteed to hit the full gen 3 speeds ?
I've actually had no issue getting LLM's to run on older graphics cards and with Llama 3.2's models being under 2GB you can spin up a baremetal LLM with Ollama without too much issue (assuming the software support is there). I would be very curious if it would be possible to get a slightly more modern AMD GPU that's Ollama supported running, especially once multi-modal support is in it could make for an interestingly sized computer vision system.
When I check the release page, they have ollama-linux-arm64. Then again, an LLM can also still be CPU bound even if they have enough GPU, so I think that might be a problem.