The title is 99% pure clickbait. There is one single mention of "M4": "I can run more displays off this single Pi 5 than a brand new M4 Pro Mac Mini." which doesn't support the broad claim of the title. That's 12 minutes I'll never get back.
Doom Eternal on a raspberry pi still hitting 30fps is insane. Even though the use case for a dgpu on a pi is limited, I think this demonstrates just how far modern ARM cpus can go on a tight power budget.
Alternately, it shows just how incompetent most game CPU coding is, that some no more complicated games can't hold 60fps on a modern desktop CPU. Meanwhile Doom Eternal can do 1000fps.
I do admire Jeff Geerling because he is always willing to learn, credit where credit is due, and most importantly, doesn't drive at 90Mph on a residential area.
I love the GPU madness series of videos - absolute fave..probably because I remember the early days of linux where getting X to run was a massive deal.
Nowadays I'm constantly impressed with Mesa, Vulkan, etc. how things 'just work' once you get the physical layer and base driver going. Just 10 years ago I'd probably spend more time battling a desktop environment instead of... doing fun stuff on Linux :D And Proton/Wine/Steam/Box86 really drives that home!
New games are big because the textures are huge. It boils down to objects not looking potato when you get close to them while playing at 4k, especially on large displays like TVs. There are some beautiful games with minimal amount of textures and mostly shader and particle effects driven, but those are few and far between, and often not demanding. Like thumper
Also, there’s lack of optimisation due to lazy development strategies that are often a result of increased cases of time crunch, which means MANY more games are being cranked out at once but the output of a game studio becomes sloppier as a result.
@@fujinshu I don't know what world you're living in where "MANY more games are being cranked out at once". Pick practically any studio, you'll see that the output is way down compared to what it was in the past.
I remember thinking that the Raspberry Pi would be the only notable piece of tech that could not run Crysis or Doom, but now I can finally put that theory to rest...
Eben stated when the Raspberry Pi 5 first lunched that a 16GB model was possible. I've been crossing my fingers that they would release that soon since then. But it looks like Forza Horizon 4 could run well on here with more RAM. I would love to have a 16GB Pi as they make great web servers.
13:49 just going by the error message it appears to be related to loading unsigned/invalidly signed firmware, or something like that. If it hasn't been fixed yet, I'm sure it will be soon.
As a workaround for not having ROCm on ARM, consider trying Vulkan acceleration for LLMs instead. Performance takes about a 25% hit in my experience, but it's something.
@@JeffGeerling you should still be able to bypass ollama and load the underlying models through code e.g. tensorflow or pytorch. Those libraries have native ARM support, though to get hardware acceleration working you may need to specify that you want to use Vulkan or OpenGL, as they won't assume that by default.
OpenCL is also available on AMD gpus via rusticl if you have an up to date version of mesa if ollama or other llms can use that (provided it works on ARM)
@@JeffGeerling you can use llama.cpp directly as Ollama also uses llama.cpp in the background. ARM is supported, as it can be used with Apple Metal. But performance wise, in my experiments, its 2-3 times slower than ROCm.
Epic Look Mum No Computer outro! Love the ARM64 PCIe experiments a lot, looking forward to the Ampere content too. Running a Snapdragon ARM64 my main machine since a year with Windows, but ARM64 it really can't be seen as a real PC alternative. I can't even run Linux on it properly, because Qualcomm isn't interested, also I couldn't add a big GPU. In general with the ARM64 stuff support is really dependent on specific interests of the big manufacturers (Nvidia: Linux + AI, totally works, Windows + Gaming.. not at all :D) your efforts and documentation really pushes things that should be working already to.. working - keep it up!
Margins on such low-mid cards are lowish because there's competition (would be lower if AMD had any mindshare). It's the high-end cards from nVidia that have no competition that give high margins, nevermind the AI stuff that has like 90% gross margin.
if you priced all of the componants on a 4090 that card would cost you less than $300 to build. the profit margines are insane, on almost every product anyone buys because there are zero regulations on how much a company can rip someone off if they will pay the price.
@@bdhale34 You're talking out of your ass. While nVidia's profit margins are probably the highest in the industry (outside of budget cards at least), that means ~50% gross margin. The 4090 BoM is a hell of a lot more than $300. The GPU die cost alone is ~$225 assuming 90% yields.
@ The tech hardware industry in particular has always seemed vulnerable to price fixing due to there only being a few suppliers and a few leading design IPs for memory, CPUs, GPUs, storage and the like. It’s not like Micron is a bakery in New York competing with a thousand others. I’m convinced that a 24 terabyte “Enterprise” magnetic drive might cost 50% more to manufacture at scale than a 5 terabyte “consumer” one now- and the absurd premium is just a worldwide price fixing cartel.
So next step - fast network switch, multiple RPis with ramdisk shared via NFS and mounted as swap. I did that years ago, shared 12GB over 100Mb/s network into 1GB RAM netbook via NFS. I was very surprised how well that worked! :)
When I tried it with NBD, I always lost the connection when it tried to swap back in, but I haven't tried NFS for swap yet. If it works reliably, that'd be a good option to deal with the memory hog that is steamwebhelper.
I did a quick Geekbench 6 compute on the M4 Max 64GB and it it outpacing the AMD Radeon Pro W7700. Pretty impressive that a 14" notebook GPU is even in the same league as a 190W dual slot desktop card.
Sheeeeesh! M4 Max is amazing-and I'm hopeful they intro an M4 Ultra at some point. I would loooove to upgrade my Studio now that my home Mac is faster than my 'work' Mac!
@@JeffGeerling What is funny is I ended up using the M1 Max 64GB laptop all the time, and much more seldomly the M1 Ultra 128GB Studio. Portability is key for me.
How far has Asahi Linux gotten now ? They got all major non-GPU hardware working fine on the m1 and m2 I believe ? And GPU progress was really good I think ?
Its never a dumb idea to make something work, its only inefficient and a bit silly if you have to buy expensive high performance new stuff that you know is going to bottleneck in your intended real world deployment. But getting a GPU working on a low power CPU, or with unusual architecture opens up so many sane usecases.
You have great taste in games. Portal and Portal 2 are about the peak of gaming experience IMO! Perfect balance of gameplay, graphics, storyline, and soundtrack, plus a really innovative and flexible game mechanic.
@@JeffGeerling I have two old iMac G4's, but they are not the widescreen/higher resolution models. If I did this, I'd have to have a widescreen iMac G4. But yeah, I'm a bit jealous of his new iMac G4-M4. I will say though, it's probably the only video Sean has made that made me want to replicate it... everything else just looks like torture, but it's entertainment for us!
Thank you for your work making this possible. I'm sure it plays some role in AMD drivers development. ANd second - for sharing "Action Retro", guy seem good, I wonder why i never stumbled on hm before.
This is quite impressive for sure. I just always find myself wanting more pcie lanes on the pi. I know that's an engineering task, but imagine having even one more available. Or a total of four, and even at gen 3 speeds, the bandwidth is tantalizing. Splitting that between nvme and a GPU feels like a dream.
Finally you did it! And we did it! I m following your videos about this pi+gpu concept. I would love to see a video with RX 6400 from amd working with a pi5. Maybe test New Vegas, Bioshock, Assassins Creed series.
Modern AAA games are big because they include huge high resolution assets like massive textures and their mipmaps, incredibly detailed models with often 6 level of detail, etc. It then has to include all the lower resolution assets for lower quality settings. Console games are often smaller because they can specifically tune the asset pool, but on PC you don't know if the person running your game is going to run it on low or ultra settings so you have to include everything to make that possible.
Would be interesting to have a future CM5 connected to a laptop formfactor GPU (MXM) on a carrier board. Some of those are pretty cheap, like a used rx 5500xt with MXM connecter can be had for 40€ over here
A reason for why Doom Eternal ran so well is that most (or all, I don't remember) of the lighting is pre-baked into textures (also why the download is massive).
I did a lot of comparisons between 1080p and 4K, and surprisingly, the performance is usually close enough I prefer 4K on my larger monitor. The CPU/bus bottleneck clears up at much lower resolutions (like 720p or less), but then things get a little more pixelated :( There's still room for improvement in the driver though-it's doing a couple tweaks that take away a little performance. And for things like OBS, there are other ways to optimize screen capture I haven't explored.
The proof of concept is great. It doesn't need to be practical, it will find it's use cases and this is why we youtube. Better dust off the 1080ti(wonder if I could get that to work). I can prepare my underground apocalypse bunker with a back-up gaming PC now.
When you replied "why not Half-Life 2 in 4k60?" I was like "errrr....", now I see the results. Impressive progress there! As someone who really likes PC games of that era anyway, the Pi 5 could be fun to use for a retro gaming rig.
In Chinese, we call this hardware combination '吕布骑狗'.🤣 Okay, to be serious, it's a really useful test. Can't wait to see Compute Module 5 with more extensibility.
I realize it's in your wheelhouse of interest, and R Pi probably encourage you by sending you stuff, but you do such a good job making things look fun and accessible. I started thinking about seriously playing around with Raspberry Pis about a year ago. My frustrations with windows and seeing if I'd be more satisfied with Linux only making me consider more, and I stumbled on your channel when doing a little research and you quickly had me excited about a number of possibilities with the Pi 5. To the point I'm thinking I should buy a R Pi 5 or two this year in the coming weeks. One for retro, one for just normal stuff and learning Linux to see what I can do with it when I don't need my gaming laptop. (I'll probably start with one and have two different storage / boot drives). I'm still hoping that they create an (semi-)official Graphics Card hat or base to make that aspect easier for mid-range rendering. But that's probably another generation out, unless the market suddenly has a massive demand for it. Although I wonder what the boards for the CM 5 will offer when it becomes available. But if there's a reasonably accessible and inexpensive way to get the Pi to emulate like a PS4 Pro for gaming (~10 year old Mid-range gaming PC), that would be the point where I couldn't ask for more to be fully satisfied. The Pi 5 is already basically there, but the GPU is definitely not.
The best thing about SBCs (whether Pi or something else) is I can feel amazing if I get something to work (it's boring getting things working when they're expected to work)... and I also don't get depressed when I completely fry one, because that's $50 or $60 and not $600 or $1000 of computer I just cooked :D
I think one reason why games are so memory-laden is better to keep up with the speed and demand of performance today. they don't compress their files as much so it is available right away. plus since today's peripherals and PC components can "dew it", developers don't bother trying to squeeze out any unnecessary bytes. so basically I'm sure there's a way they can like cut their code in half and save on memory, but because computers are so powerful and people have huge hard drive capacity anyway, they don't bother. plus I'm sure manufacturers of them are happy when you upgrade your storage to fit the next never ending games that are released
This isn't just fun to do, but a peek into the future! What would blow my mind would be if the asahi linux team could develop a thunderbolt driver, we could be using nvidia/amd gpus on macs again while running linux lol
Really excited about this, it would be great to see this on another more powerful ARM SBC, but I am sure that would take way too much effort to get it working on boards with way smaller Userbases. But this on an rk3588 with 16GB of RAM would be very interesting.
Definitely-I once tried a year ago, haven't since. But one of the big issues is trying to get newer Linux kernels on these boards, since a lot of driver fixups happen between like 5.10 to 6.1, and 6.1 to 6.6, etc. On the Pi, it's official Pi OS stays pretty close to modern LTS releases, so it's been easier to maintain patches.
I will be super excited about this if big Nvidea CUDA cards for Ollama and local LLMs become plausible. Thanks for letting us know right at the front of the video, but I still watched to the end. :)
I have two hopes. One, your GPU patches gets added to the main Raspberry Pi Kernel. Two, Rockchip for their 35XX have this PCI-E development get better to people can use GPUs on those things.
Speaking of display outputs, having a gaming laptop with dGPU kind of saved my bacon in theatre tech. We needed to run 4 separate projectors simultaneously at one point in our play, and most laptops with iGPU couldn’t do it and keep the internal display. Using both the iGPU (USB-C alt mode to HDMI converters) and dGPU (the latter assigned to the internal display via mux), I was able to pull it off, with room for one more (the HDMI is connected to the dGPU). I found that hyper specific niche case that a Macbook cannot touch. 😝
6:37 it's cool to see what theraspberry pi can become. A small piece of electronics that you just to plug it into your TV and play your games. I currently do this with my laptop, but it's kind of large and bulky, and in the way.
I've been saying it for years, if we get x86/ARM translation going well ARM portables will dominate, the current SD chips are beasts, they just need the PC library for take over the handheld space, maybe ARM+AMD is also a good combo, hope in some years the Steam Deck 3 run games at 1080p for like 6 hours
i am just a user and i think that this video is big for dedicated devices for different uses. separate gaming on small size (I do 720 mostly fullscreen on 1080 displays or windowed in 4k tvs), separate local ai assistant later. separate home security later. I am thinking for a long time to the radxa x4 with the intel n100 to run windows to have an easy to use and configure security cameras software to capture, store and have it available, do backup of files and folders i want with a 3rd party software, and have a local Steam cache to pull games from there, instead of the internet, to my main pc. and rpi and gpus open up new potentials for llm to the mix
Thanks for sharing this, now I need to go to make a cluster out of as many Raspberry Pi 5 I can get! And add some AMD Top Tires GPUs rack to the cluster.
I can just picture a package arriving at Jeff's office while he was filming that surfshark ad spot. "Hello, Package for Jeff? Uhhh...Yes. You can put it down over there please, I don't have hands at the moment".
fun. there are a lot of cool possibilities for retrogaming on a Pi5 with an external GPU. we can more easily connect to an external VGA monitor, or even a downscaler if you want to display on 15kHz display
For LLMs, since you can run Vulkan, you should be able to get llama.cpp or MLC with the Vulkan backends working. While compute (prompt processing) is much slower than ROCm, you should be able to get pretty good token generation speed.
Ha, I actually just bought it this week, because I've been carrying that PC from place to place all the time, and finally tweaked my back doing it. It's a Eureka PC cart I found on Amazon.
Hey Jeff, check out Portal Stories: Mel, and Portal Reloaded. As a portal fan, these were extremely enjoyable! Side note: I am stunned at how many games you got running. Even ignoring the mind boggling fact you've got a dGPU behaving on a pi now, I did not expect games to run on arm on linux like that. Very cool stuff!
With game size, think of it like this: There's 10 times as many objects with 10 times as much resolution, that's 100 times as much data to store. Each quality of the game is a multiplier and they're trying to cram more into every category at once, so it's not linear without major optimizations.
I have a idea use a mobile amd card (usually used for laptop gpus) then connect it to a raspberry pi and put it in a small case and install bazzite or something that's similar to steam os to make a steam machine or you can use a normal low profile amd gpu
You may be able to connect more monitors by streaming to other devices, and also vnc...(...aaand you already thought about streaming. (LOL!, commenting before the end of the video))
Awesome as always.. thanks! as for the lack of RAM? just "download more RAM!" (and if you remember *that* reference - and false advertising - yer as old as me!)
I'd love to do this. Just did an upgrade/hand-me-down on 2 house pcs and now have a leftover GTX750 Ti SC. It's PCIe 3.0 and only runs at 60 TDP but still outperforms the VC VII by about 2000%! I have a leftover 400W sitting around now too. If only someone had already figured it out....
Surely you mean your GPU has a better Raspberry Pi than the M4 Pro does ?
Still impressive arm support here ... The openess of rasp5 Is something average Apple simp Will never understand ...
The title is 99% pure clickbait. There is one single mention of "M4":
"I can run more displays off this single Pi 5 than a brand new M4 Pro Mac Mini."
which doesn't support the broad claim of the title. That's 12 minutes I'll never get back.
🤣
@@-danR This title is not clickbait, it's video about running external GPU on RPi 5 that is more powerful than M4 integrated GPU.
@@-danR No sense of humor, huh?
All I wanna know is if I can play Jeff Geerling videos in 4K.
Doom Eternal on a raspberry pi still hitting 30fps is insane. Even though the use case for a dgpu on a pi is limited, I think this demonstrates just how far modern ARM cpus can go on a tight power budget.
imagine if it was a Snappy X Elite
@@123Andersonev I soooooo wish Qualcomm would see 'real' PCIe expansion as something worth adding. The Dev Kit came close but didn't deliver there.
@@JeffGeerling Full PCIe for ARM would be a game changer, we might even see more development for faster Raspberry Pi compute modules.
@@AndrewAHayes Doesn’t Ampere have that? Not at all an SBC, but still ARM.
Alternately, it shows just how incompetent most game CPU coding is, that some no more complicated games can't hold 60fps on a modern desktop CPU. Meanwhile Doom Eternal can do 1000fps.
babe come quick, they got GPUs working on the Raspberry Pi 5
This might allow Nintendo switch,Xbox 360, ps 3 emulation on a pi might still need a pi 6 with better cpu performance but we're close
4:37 Linus would be proud of that segue.
No way 😂😂😂😂
Nobody cares.
@@tablettablete186 He means that guy from LTT, not the one from the LF.
I do admire Jeff Geerling because he is always willing to learn, credit where credit is due, and most importantly, doesn't drive at 90Mph on a residential area.
Yet most people would still watch and believe that person who drives at 90mph in a residential area simply because he has millions of subs.
He doesn’t sell $50 jpgs either 😂
But he did to run 4K on a Pi GPU
i don’t get the reference
@@TheEmptyHoliness It’s about Marques Brownlee.
Shoutout to Action Retro, LMNC music in the background. What a time to be alive :D
"Super easy to install steam, barely an inconvenience". I see what you did there Jeff.
I missed that pitch meeting moment.
Wow wow wow!
He said the thing from the other thing!
@@Reeonimus He did it that way because I needed him to do it so the movie could happen. Besides, that's the way physics works.
Using six monitors on a Pi5 is tight! Cheers, daveyb
I love the GPU madness series of videos - absolute fave..probably because I remember the early days of linux where getting X to run was a massive deal.
Nowadays I'm constantly impressed with Mesa, Vulkan, etc. how things 'just work' once you get the physical layer and base driver going. Just 10 years ago I'd probably spend more time battling a desktop environment instead of... doing fun stuff on Linux :D
And Proton/Wine/Steam/Box86 really drives that home!
New games are big because the textures are huge. It boils down to objects not looking potato when you get close to them while playing at 4k, especially on large displays like TVs. There are some beautiful games with minimal amount of textures and mostly shader and particle effects driven, but those are few and far between, and often not demanding. Like thumper
Also, there’s lack of optimisation due to lazy development strategies that are often a result of increased cases of time crunch, which means MANY more games are being cranked out at once but the output of a game studio becomes sloppier as a result.
@@fujinshui ❤ uncompressed 4k textures
It goes beyond that also though. The current triple AAA dev cycle doesn't prioritize compression or optimization as a huge priority.
@@fujinshu I don't know what world you're living in where "MANY more games are being cranked out at once". Pick practically any studio, you'll see that the output is way down compared to what it was in the past.
Even still they don't optimize for shit. Just look at how much memory Hitman 2 & 3 got down to
Jensen's Crysis comment had some serious "How do you do, fellow kids" energy
Only gamers would get that joke!
I remember thinking that the Raspberry Pi would be the only notable piece of tech that could not run Crysis or Doom, but now I can finally put that theory to rest...
I feel like they can all run doom (90s doom at least)
Cool that u used the music of "look mum no computer". U should do that more!
Looooove his channel, his vibe, his music :D
Just want to go visit that museum next time I'm in the UK.
3:30 Wow wow wow Wow
Oh wow.
Ryan George references are tight!
@@JeffGeerlingwow wow wow wow
Eben stated when the Raspberry Pi 5 first lunched that a 16GB model was possible. I've been crossing my fingers that they would release that soon since then. But it looks like Forza Horizon 4 could run well on here with more RAM. I would love to have a 16GB Pi as they make great web servers.
16GB would be great for ZFS as well.
Can you get away with using Ram on a GPU?
13:49 just going by the error message it appears to be related to loading unsigned/invalidly signed firmware, or something like that. If it hasn't been fixed yet, I'm sure it will be soon.
As a workaround for not having ROCm on ARM, consider trying Vulkan acceleration for LLMs instead. Performance takes about a 25% hit in my experience, but it's something.
So far I can't get the Ollama stuff working with Vulkan because it's X86 only, but there may be other ways... still tinkering with that.
@@JeffGeerling you should still be able to bypass ollama and load the underlying models through code e.g. tensorflow or pytorch. Those libraries have native ARM support, though to get hardware acceleration working you may need to specify that you want to use Vulkan or OpenGL, as they won't assume that by default.
OpenCL is also available on AMD gpus via rusticl if you have an up to date version of mesa if ollama or other llms can use that (provided it works on ARM)
@@JeffGeerling you can use llama.cpp directly as Ollama also uses llama.cpp in the background. ARM is supported, as it can be used with Apple Metal. But performance wise, in my experiments, its 2-3 times slower than ROCm.
Yet again pushing boundaries. Kudos. There are rumours of CM5 with 16gb, may make a lot more games playable on pi egpu setup.
I sure hope so!
Epic Look Mum No Computer outro! Love the ARM64 PCIe experiments a lot, looking forward to the Ampere content too.
Running a Snapdragon ARM64 my main machine since a year with Windows, but ARM64 it really can't be seen as a real PC alternative. I can't even run Linux on it properly, because Qualcomm isn't interested, also I couldn't add a big GPU. In general with the ARM64 stuff support is really dependent on specific interests of the big manufacturers (Nvidia: Linux + AI, totally works, Windows + Gaming.. not at all :D) your efforts and documentation really pushes things that should be working already to.. working - keep it up!
This makes me wonder what the margins are on $300 and $400 GPUs.
I just wish there were more budget GPUs available these days :(
Margins on such low-mid cards are lowish because there's competition (would be lower if AMD had any mindshare). It's the high-end cards from nVidia that have no competition that give high margins, nevermind the AI stuff that has like 90% gross margin.
if you priced all of the componants on a 4090 that card would cost you less than $300 to build. the profit margines are insane, on almost every product anyone buys because there are zero regulations on how much a company can rip someone off if they will pay the price.
@@bdhale34 You're talking out of your ass. While nVidia's profit margins are probably the highest in the industry (outside of budget cards at least), that means ~50% gross margin. The 4090 BoM is a hell of a lot more than $300. The GPU die cost alone is ~$225 assuming 90% yields.
@ The tech hardware industry in particular has always seemed vulnerable to price fixing due to there only being a few suppliers and a few leading design IPs for memory, CPUs, GPUs, storage and the like. It’s not like Micron is a bakery in New York competing with a thousand others.
I’m convinced that a 24 terabyte “Enterprise” magnetic drive might cost 50% more to manufacture at scale than a 5 terabyte “consumer” one now- and the absurd premium is just a worldwide price fixing cartel.
So next step - fast network switch, multiple RPis with ramdisk shared via NFS and mounted as swap. I did that years ago, shared 12GB over 100Mb/s network into 1GB RAM netbook via NFS. I was very surprised how well that worked! :)
I mean, I shared 12GB of RAM from desktop PC to Intel Atom netbook and just browsed web, but it might work for games too.
When I tried it with NBD, I always lost the connection when it tried to swap back in, but I haven't tried NFS for swap yet. If it works reliably, that'd be a good option to deal with the memory hog that is steamwebhelper.
SSH -X bigiron firefox
I did a quick Geekbench 6 compute on the M4 Max 64GB and it it outpacing the AMD Radeon Pro W7700. Pretty impressive that a 14" notebook GPU is even in the same league as a 190W dual slot desktop card.
Sheeeeesh! M4 Max is amazing-and I'm hopeful they intro an M4 Ultra at some point. I would loooove to upgrade my Studio now that my home Mac is faster than my 'work' Mac!
@@JeffGeerling What is funny is I ended up using the M1 Max 64GB laptop all the time, and much more seldomly the M1 Ultra 128GB Studio. Portability is key for me.
How far has Asahi Linux gotten now ? They got all major non-GPU hardware working fine on the m1 and m2 I believe ? And GPU progress was really good I think ?
@@ServeTheHomeVideo True, you put in a lot more miles than I do!
@@autohmae It's great on M1/M2. Almost indistinguishable from x86 Fedora. My M3 Pro however cannot run it yet.
Its never a dumb idea to make something work, its only inefficient and a bit silly if you have to buy expensive high performance new stuff that you know is going to bottleneck in your intended real world deployment. But getting a GPU working on a low power CPU, or with unusual architecture opens up so many sane usecases.
I have to say your silliest content are my fave
You have great taste in games. Portal and Portal 2 are about the peak of gaming experience IMO! Perfect balance of gameplay, graphics, storyline, and soundtrack, plus a really innovative and flexible game mechanic.
0:26 the Team Rocket card :D
:D
♥
👋
But if you watch his latest iMac G4-M4 video though, he actually made something useful and good looking too... I think somethings wrong!
@@Toby_Q Haha I know... at the end of that video I was like ...and I want one now.
@@JeffGeerling I have two old iMac G4's, but they are not the widescreen/higher resolution models. If I did this, I'd have to have a widescreen iMac G4. But yeah, I'm a bit jealous of his new iMac G4-M4. I will say though, it's probably the only video Sean has made that made me want to replicate it... everything else just looks like torture, but it's entertainment for us!
Thank you for your work making this possible. I'm sure it plays some role in AMD drivers development.
ANd second - for sharing "Action Retro", guy seem good, I wonder why i never stumbled on hm before.
Thats quite the accomplishment Jeff. Excellent work dude. 👍
Kudos for the music by Look Mum, No Computer!
Looking forward to the Pi6 with Thunderbolt 4.
Haha!
With the rate at which Broadcom/Pi can adopt new tech, it'd probably be Thunderbolt 2 (which, ironically, would still be pretty useful :D)
Once the drivers get stable, a pi + a dedicated gpu could be genuinely a good enough combo for casual usage, and even games by the looks of it
Jeff is the definition of a 'madlad'
This is quite impressive for sure. I just always find myself wanting more pcie lanes on the pi. I know that's an engineering task, but imagine having even one more available. Or a total of four, and even at gen 3 speeds, the bandwidth is tantalizing. Splitting that between nvme and a GPU feels like a dream.
Or dual 10 GbE Ethernet :D
Finally you did it! And we did it! I m following your videos about this pi+gpu concept. I would love to see a video with RX 6400 from amd working with a pi5. Maybe test New Vegas, Bioshock, Assassins Creed series.
Modern AAA games are big because they include huge high resolution assets like massive textures and their mipmaps, incredibly detailed models with often 6 level of detail, etc. It then has to include all the lower resolution assets for lower quality settings. Console games are often smaller because they can specifically tune the asset pool, but on PC you don't know if the person running your game is going to run it on low or ultra settings so you have to include everything to make that possible.
They should just target 480p, problem solved!
But do they need to include everything? Couldn't they ship low-medium textures to save space, then have the ultra resolution textures as free DLC?
thank you jeff for paving the way for the future on arm. I can definitely see valve making the steam deck 3/4 a full arm chip.
Absolute mad lad.
BTW your Ansible guides helped me a lot landing a job so thank you good sir.
Would be interesting to have a future CM5 connected to a laptop formfactor GPU (MXM) on a carrier board. Some of those are pretty cheap, like a used rx 5500xt with MXM connecter can be had for 40€ over here
Such a flex! For piheads everywhere, this is great stuff
A reason for why Doom Eternal ran so well is that most (or all, I don't remember) of the lighting is pre-baked into textures (also why the download is massive).
Probably wont see it till next gen, but id love to see 16gb and 32gb models of the pi.
8:48 thats an xbox one s or newer controller
True... no glowing green X orb!
Thanks for anticipating the question about LLMs and ROCm!
“Super easy barely an inconvenience”
Bottlenecking your graphics card with a potato is always a good policy.
Interested in seeing the perf of this setup in 1080p 😊
I did a lot of comparisons between 1080p and 4K, and surprisingly, the performance is usually close enough I prefer 4K on my larger monitor. The CPU/bus bottleneck clears up at much lower resolutions (like 720p or less), but then things get a little more pixelated :(
There's still room for improvement in the driver though-it's doing a couple tweaks that take away a little performance. And for things like OBS, there are other ways to optimize screen capture I haven't explored.
I don't even own a 4k TV yet. But nobody is complaining.
The proof of concept is great. It doesn't need to be practical, it will find it's use cases and this is why we youtube.
Better dust off the 1080ti(wonder if I could get that to work). I can prepare my underground apocalypse bunker with a back-up gaming PC now.
They finally got steamlink working for raspberry pi 5 😭 amazing!
When you replied "why not Half-Life 2 in 4k60?" I was like "errrr....", now I see the results.
Impressive progress there! As someone who really likes PC games of that era anyway, the Pi 5 could be fun to use for a retro gaming rig.
wow its really impressive considering most of these games are running under wine which also adds some overhead
Yeah, and that the overhead is mostly more RAM usage and a bit CPU makes it worse
I had to double check that it wasn't April 1st....incredible work...
In Chinese, we call this hardware combination '吕布骑狗'.🤣
Okay, to be serious, it's a really useful test. Can't wait to see Compute Module 5 with more extensibility.
Thanks for the video, it was news to me how many commercial games are available with natively compiled arm binaries already
I realize it's in your wheelhouse of interest, and R Pi probably encourage you by sending you stuff, but you do such a good job making things look fun and accessible.
I started thinking about seriously playing around with Raspberry Pis about a year ago. My frustrations with windows and seeing if I'd be more satisfied with Linux only making me consider more, and I stumbled on your channel when doing a little research and you quickly had me excited about a number of possibilities with the Pi 5.
To the point I'm thinking I should buy a R Pi 5 or two this year in the coming weeks. One for retro, one for just normal stuff and learning Linux to see what I can do with it when I don't need my gaming laptop. (I'll probably start with one and have two different storage / boot drives).
I'm still hoping that they create an (semi-)official Graphics Card hat or base to make that aspect easier for mid-range rendering. But that's probably another generation out, unless the market suddenly has a massive demand for it. Although I wonder what the boards for the CM 5 will offer when it becomes available.
But if there's a reasonably accessible and inexpensive way to get the Pi to emulate like a PS4 Pro for gaming (~10 year old Mid-range gaming PC), that would be the point where I couldn't ask for more to be fully satisfied. The Pi 5 is already basically there, but the GPU is definitely not.
The best thing about SBCs (whether Pi or something else) is I can feel amazing if I get something to work (it's boring getting things working when they're expected to work)... and I also don't get depressed when I completely fry one, because that's $50 or $60 and not $600 or $1000 of computer I just cooked :D
I think one reason why games are so memory-laden is better to keep up with the speed and demand of performance today. they don't compress their files as much so it is available right away. plus since today's peripherals and PC components can "dew it", developers don't bother trying to squeeze out any unnecessary bytes. so basically I'm sure there's a way they can like cut their code in half and save on memory, but because computers are so powerful and people have huge hard drive capacity anyway, they don't bother. plus I'm sure manufacturers of them are happy when you upgrade your storage to fit the next never ending games that are released
This isn't just fun to do, but a peek into the future! What would blow my mind would be if the asahi linux team could develop a thunderbolt driver, we could be using nvidia/amd gpus on macs again while running linux lol
My hope, tbh!
Amazing this would be great for emulation and possibly Tensorflow in the future.
Really excited about this, it would be great to see this on another more powerful ARM SBC, but I am sure that would take way too much effort to get it working on boards with way smaller Userbases. But this on an rk3588 with 16GB of RAM would be very interesting.
Definitely-I once tried a year ago, haven't since. But one of the big issues is trying to get newer Linux kernels on these boards, since a lot of driver fixups happen between like 5.10 to 6.1, and 6.1 to 6.6, etc.
On the Pi, it's official Pi OS stays pretty close to modern LTS releases, so it's been easier to maintain patches.
thank you for being yourself, for your content and being a role model in this mad world!
I will be super excited about this if big Nvidea CUDA cards for Ollama and local LLMs become plausible. Thanks for letting us know right at the front of the video, but I still watched to the end. :)
I have two hopes. One, your GPU patches gets added to the main Raspberry Pi Kernel. Two, Rockchip for their 35XX have this PCI-E development get better to people can use GPUs on those things.
"Mom i want a gaming pc" "but we have a gaming pc at home" The gaming pc at home:
I appreciated the "barely an inconvenience" reference 🙂
Speaking of display outputs, having a gaming laptop with dGPU kind of saved my bacon in theatre tech. We needed to run 4 separate projectors simultaneously at one point in our play, and most laptops with iGPU couldn’t do it and keep the internal display.
Using both the iGPU (USB-C alt mode to HDMI converters) and dGPU (the latter assigned to the internal display via mux), I was able to pull it off, with room for one more (the HDMI is connected to the dGPU).
I found that hyper specific niche case that a Macbook cannot touch. 😝
There's always a niche use case for even the wildest of setups!
As an owner of a Pi and an old rx 580, I can only wonder what I could do with the two combined.
*Super easy, barely an inconvenience* I see what you did there!
6:37 it's cool to see what theraspberry pi can become. A small piece of electronics that you just to plug it into your TV and play your games. I currently do this with my laptop, but it's kind of large and bulky, and in the way.
I've been saying it for years, if we get x86/ARM translation going well ARM portables will dominate, the current SD chips are beasts, they just need the PC library for take over the handheld space, maybe ARM+AMD is also a good combo, hope in some years the Steam Deck 3 run games at 1080p for like 6 hours
The Radeons supports DisplayPort Multistream. So with daisy chained monitors and or a DisplayPort MSt-Hub you could add at least two more monitors.
Oooh... I wonder if Linux supports that correctly too?
Phoronix posted an article in 2021 that DP 2.0 support, including MST, was merged into kernel 5.16. So it _should_ work.
I do have a few amd gpus laying around... and I do just need a PC to do 1080p recording and streaming....... Thanks Jeff! Now I know what to do ;)
i am just a user and i think that this video is big for dedicated devices for different uses. separate gaming on small size (I do 720 mostly fullscreen on 1080 displays or windowed in 4k tvs), separate local ai assistant later. separate home security later. I am thinking for a long time to the radxa x4 with the intel n100 to run windows to have an easy to use and configure security cameras software to capture, store and have it available, do backup of files and folders i want with a 3rd party software, and have a local Steam cache to pull games from there, instead of the internet, to my main pc. and rpi and gpus open up new potentials for llm to the mix
Thanks for sharing this, now I need to go to make a cluster out of as many Raspberry Pi 5 I can get! And add some AMD Top Tires GPUs rack to the cluster.
I knew this would improve fast but not this fast! Pi support is unrivaled.
I can just picture a package arriving at Jeff's office while he was filming that surfshark ad spot. "Hello, Package for Jeff? Uhhh...Yes. You can put it down over there please, I don't have hands at the moment".
It's just a challenge figuring out how to do everyday things with shark-hands!
This is really cool. I would love to see a Arm-based steamdeck, but thought it is only a place in my dreams, but this brings me hope.
fun. there are a lot of cool possibilities for retrogaming on a Pi5 with an external GPU. we can more easily connect to an external VGA monitor, or even a downscaler if you want to display on 15kHz display
I'm fairly sure you can run LLMs and other models on this with GPU support using something which has it's own userspace like tinygrad !
@jeffgerling Jeff keep on rocking cos I learn a lot from you. You rock from NZ
For LLMs, since you can run Vulkan, you should be able to get llama.cpp or MLC with the Vulkan backends working. While compute (prompt processing) is much slower than ROCm, you should be able to get pretty good token generation speed.
How does the saying go... "This $#!T is banana's" ... LOL
B-A-N-A-N-A-S!
@@JeffGeerling Nice! hahahaha
Quick question, what is that trolley you are using for your gaming PC? I need one!
Ha, I actually just bought it this week, because I've been carrying that PC from place to place all the time, and finally tweaked my back doing it. It's a Eureka PC cart I found on Amazon.
When the device is cheaper than the game itself.
This makes me want to try getting Steam to work on Ampere and my M1 Pro MBP running NixOS.
Wait, can macOS run NixOS?!?!
I thought only Asahi could run so far!
Don't you mean your GPU has a Raspberry pi? Lol! Great video!
Hey Jeff, check out Portal Stories: Mel, and Portal Reloaded. As a portal fan, these were extremely enjoyable!
Side note: I am stunned at how many games you got running. Even ignoring the mind boggling fact you've got a dGPU behaving on a pi now, I did not expect games to run on arm on linux like that.
Very cool stuff!
Oh no... I have avoided Reloaded for years... don't tempt me! lol
@JeffGeerling The time has come. Pi: Reloaded running Portal: Reloaded
With game size, think of it like this: There's 10 times as many objects with 10 times as much resolution, that's 100 times as much data to store. Each quality of the game is a multiplier and they're trying to cram more into every category at once, so it's not linear without major optimizations.
A super-niche use case for an external GPU + Pi5 may be a tiny FFmpeg render server.
I have a idea use a mobile amd card (usually used for laptop gpus) then connect it to a raspberry pi and put it in a small case and install bazzite or something that's similar to steam os to make a steam machine or you can use a normal low profile amd gpu
Very impressive progress on this! I guess in no small part down to your own personal obsession! 🙂
Nice vid mate
You may be able to connect more monitors by streaming to other devices, and also vnc...(...aaand you already thought about streaming. (LOL!, commenting before the end of the video))
Come for the FPS, stay for the bloopers.
How do we even get steam and steam link going on pi 5? I tried so many different installations when I got my 5 but could never get anything going.
`sudo apt install -y steamlink` - they just updated it this week :)
@@JeffGeerling Sick! Thanks for the reply (and the vid)!
Awesome as always.. thanks! as for the lack of RAM? just "download more RAM!" (and if you remember *that* reference - and false advertising - yer as old as me!)
Oh yes, I love the Xbox series 60 controller
I'd love to do this. Just did an upgrade/hand-me-down on 2 house pcs and now have a leftover GTX750 Ti SC.
It's PCIe 3.0 and only runs at 60 TDP but still outperforms the VC VII by about 2000%! I have a leftover 400W sitting around now too.
If only someone had already figured it out....
Hey Jeff. Do you plan to try same setup with RK3588 boards? They often have full PCIE gen3 x4 and twice the core number. Might work better...
14:49 if he is Jeff Geerling till next time WHAT IS HE AFTER NEXT TIME
He will become Geff Jeerling
I read "I'm in danger" like Ralph from the Simpson's and I think that's hilarous