Nice to know - even 5 years later Pi 4/5 are $50+ Steam link are nowhere to buy :(

I would love to see how it performs under Linux Mint for example.

14:07 ... @CraftComputing ;)

The only "compact" eGPU is from GPD with PSU build-in can be a great thing for handheld... Everythink with external PSU can be trouble for portable gaming but GPD can be a solution. If in future they will manage high temps, noise and heat thanks to better/cooler GPUs with 60-100W range will be really great even with portable solutions when You can find some ac plug near by...

My use case: already have a handheld and I want to wall mount an egpu behind the tv for 4k gaming and reduce wear.

I always think of 22 Jump Street when he says "I am Jeff"

this was awesome watching 2 completely different priced GPus going head to head and being matched like wtf.

Your ram speed and timings bro

What dac cable was that and are they compatible with cheaply Amazon 10g sfp plus switches

How about CSGO? Is it playable?

Rubber gasket has a pre-cut slit.

I have the hp pro sff 400 g9 with i7 12700 i got this and worked great

After six months the Liquid cooling system is overheating the CPU. To get support is almost impossible.

Thanks for the video.

way back then in early 2000's 2010's, when I was younger I'd thinker quite a bit with virtual machines, trying new different systems, I flirted quite a lot with linux, and since gaming wasn't there at all yet, I wouldnt just leave windows and distro hop, as fun as it was I couldn't really do too much, it's crazy what you can do with virtualization nowadays!

... Can you change the title of the video to specify "retrogaming" instead of just "gaming" ... This click bait was a massive waste of my time...

You can't adjust sceen resolution ? the Legion Go you can adjust on the menu

I'd like to see the AV1 transcoding performance in Plex.

THis is the kind of video you get when you have somebody who wants server hardware but doesnt have a server job to give it.

Any chance you'd try to create a DIY 8xP40 Server. I'm using a 4xP40 Server with an AMD7551p and 512gb of Ram for LLM Inference. Now that Llama 400b is upcoming I am wondering if I still could get some decent performance at 4bit with 8 or even 12 P40ties. Vram is just insanely expensive. Is there a better enterprise GPU than the P40 with the same amount or more Vram and at least the same performance, which does not cost 10x the price of a P40? 😅

maybe next you should invest into a vacuum and air compressor

I feel the server cpu was really the limiting factor more than anything... 2.3 GHz isn't great 😢

Im not sure you have made a valuable review. I feel that what you should have been comparing are retail eGPUs with 'build it yourself' eGPUs where you fit a desktop GPU to an eGPU housing. The reason to have a handheld gaming device and a desktop gaming device are completely different. You might not be able to afford BOTH a handheld device and a desktop device but you are willing to compromise. When you are at home and want a better performance, dock and eGPU, but when travelling/in bed you use just the handheld. I dont think comparing a handheld device with a desktop device is fair. Its like comparing a motor bike to a 4x4 truck.

Basically, unless you're willing and capable enough to potentially grenade a very expensive graphics cards through power rail mods and a firmware hack, you're just as well off getting an RTX 4060 LP and using the spare cash elsewhere. Like, is it theoretically possible to make the RTX 4000 Ada SFF run significantly faster? Yes, but it's very risky and you'll probably hit thermal limits pretty quickly once you add an external power rail and mod the firmware to utilize it.

Noticing that most of these old compute cards are cheap due to new AI services not supporting the older cuda version. Triton server for example supports cuda 6 and newer (going off memory here), and the tesla m40 24gb is cuda 5.2 . Did learn this the hard way too since nvidia docs weren’t clear, so wasted money on a card

there are alot of them... but we do not post them, have many - trying to make things easy for the common man... you - on the other hand - try to make this difficult, if you combine common sense - and a little google everyone will make this work,,,

I'd love to see both these GPU compared for 3D CAD workload.

Is that a Radeon VII in the background on your wall? Still rockin' one in my desktop after all these years.

I was looking at Techpowerup, the NVIDIA P102-100 is a mining gpu that has performance of a 1080 ti for under 50$ although it has 5 GB vram. I would love to see you test it. edit: the Nvidia CMP 100-210 has better performance and 16 gb vram for 200$, it is based on volta.

In short, gaming card is overall better at gaming, big suprise

Thanks for the video! Looking forward to the pascal results. Tesla P4 is a powerhouse of a LP gpu

Good experiment setup. Love these proper benchmarks

I would appreciate it if you added some highlights to your data grids when you call out specific results.

@craftcomputing Do you have an updated version of this? My passthrough was working but now it's not. Everything is still setup the same...

Using CoffeeTime I have been able to use a E3 1230 v5 in my 1151 mb, with modified bios. 4c and 8t for $30 isnt too shabby:)

the K80 was the biggest waste of money for me, all gpus are double the price in europe and the P40 while being 2.5 times more expensive has like 200% more perofrmance. also the K80 does horrible in 1080p and is pretty laggy and the mehtod of using the K80 for cloud gaming on this channel only allows for 1080p.

I virtualized my Pc using hyper-V in one of your previous videos, but I am stuck running at 1080p on the hyper-V windows 11pc i am using via Parsec. How can i increase the host resolution to say 1440p or 4k ? this doesn't seem possible ;(

Love your videos i got this server and some tesla p40s finaly got everything running is there anyway to do 144hz parsec streaming i cant find anything on it other than a dummy plug for consumer cards is there any other option since you cant plug anything into the tesla

The 7840U has a higher multi-core score in Passmark than the Xeon-E5-2697-v4. In single core it's outright humiliation.

Maxwell is not dead yet, it has D3D 12_1 support, and the lattest drivers still support them. U could still run fine any modern game with a GTX 980 Ti.

How many Patriot burst SSD's you had die on you yet? (I've had a new one last only 2 days!)

You should test the tesla T4 and A2, I really want to see those low profile single slot cards

Saving the planet from e-waste, by streaming games .. Bravo!!!

I like watching these videos for educational purposes... however the concept of streaming games is one that I will never ever support. I mean i like to watch how they setup the servers and all that, but 100% of the time I would rather run the game locally rather than stream it. Try playing a ranked dota 2 match through a streaming client......... see how well you do...... ;) Anyways keep up the videos!

Few things to mention: 1. You should probably measure GPU and CPU utilisation, and CPU utilisation per-thread. Then it should be more clear if your low 0.1%s are because of CPU or GPU. 2. CPU is indeed way underpowered. Modern games can easily use 6 cores or more and therefore it is not a good idea of giving a VM less than that. I'm not sure why not to temporary give a VM way overpowered GPU and find sweet spot for the CPU in terms of cores/threads per VM, and then use that as a base.But overall even 3.6 GHz boost clock might not be enough. 3. Another important (for dual-VM tests) thing is PCIe locality. You should always ensure that you allocate VMs to the same CPU that have the PCIe Root Complex where your GPU plugged in. Otherwise you might expereice, let's say, not so great performance scaling because your other VM will need to go via UPI to another socket and you can easily saturate the link.

Definitely interested to see the Pascal version, as I happen to have one laying around. Not something I'd want to do personally, but it's fun to know I COULD... The question is, have you ever tried butter in your coffee in place of cream?

With these being so cheap I am wondering if just using them for transcoding would be useful.

Unfortunately Kepler is the most recent Nvidia chipset still supported by Apple.

Aside from gaming, Tesla P40s make for a cheap, though slow, but still useable alternative for hosting LLMs. 24GB VRAM helps.

I've been watching all these videos for years, build my own xeon proxmox system, trying EVERY SINGLE recipe there was, for proxmox, LXC, docker (hell, you can even run windows in docker! didn't believe it till i tried it), and all I can say is: this is all great, as long as your time is FREE! If you plan to do any meaningful work on similar thing, or have it as your homelab setup - better buy something new and have fun with your kids, girlfriend or whatever. Here are realities: 1) you will spend a lot of time for something that works in youtubers case, but "may" work in your case, 2) old equipment you get will be either unsupported or with close to its physical end (remember, anything that has capacitor will slowly, over a decade - loose properties). And then when you are sick of it - who is going to buy it from you? 3) you cannot use this for any GPU intensive work other then minimum AI, mining (long time not practical) and games if you get your virtual display somehow work in your particular case. Try open source virtual displays, it is a lottery. 4) if you go by route of having virtual GPU in proxmox, you are depending on sketchy megaupload NVIDIA drivers you cannot trust, since you have no idea who messed with that, simply because, well, nvidia. 5) if you go by route of having proxmox > windows host > windows VMs, this may work - you will have miltiple vms sharing gpu, but by the time you check loss in performance, you are almost as if you can build the same bare metal system. Here I talk nvme performance, samba trough many layers, and then case of unexpected windows update that will mess you to the start. You will also discover that you actually may be needing some VPN, and that would be another pain. 6) ultimately, once you're done, you can say: "now i've learned something!", but this knowledge is short lived - it is useful just a year or so until next braking version of something is out, or HW update, when you will realize - you do not remember any more some super hackish patch, or particular kernel compile command, or, for any reason - things no longer work if just a different GPU is on another machine, even if you've documented it all. But all that fades in simple, undeniable truth: you will waste so much more time, the time you could spent with your loved ones, just to have an illusion you've saved a few bucks. And then just a few years later - all of this is already obsolete anyway.