1:48
Remember our Zotac 4090? It's getting an upgrade. It's getting It's going from 24 to 48 GB of memory. And uh it's not going to look like this anymore. You see, because when you upgrade a GPU, really you just take the GPU and the memory off of it. So, I'm going to have a naked Zotac PCB that doesn't have the important parts because you're going to go from 24 to 48 gigs. Nvidia doesn't like that. But, upgrade. Good upgrade, right? Actually the cheapest path to a 48-gig GPU that's this fast, which is absolutely absurd.
Oh, I hope this turns out okay. It's back. This is the Zotac. The giant 4090 Zotac GPU has come back to me in this box. Boom. There it is. Remove film before use. Now it's a two-slot blower-style card. What's happened? So, the upgrade process takes my chips from my board and puts them on a new board. This is really good. Also, this is probably going to be the kind of thing that gets Nvidia's knickers in a twist, but is also the kind of thing that we should have under consumer protection laws. Those are my chips. I can re-implement them however I see fit.
Before you think that is not a right that we should have, I will I No, I disagree. Completely disagree. First sale doctrine, but also a lot of established case law uh going back to the aviation industry. The Wright brothers were real handsy about having invented the aviation industry. It's like, "Whoa, you want to and then but it's just not the right not the right attitude to have." So, with this upgrade, this GPU is still a 4090, but now with 48 gigs of VRAM, it is a much more capable. This upgrade cost me I think it was about $1,600 or $1,700 or something like that. It's labeled with my name. That's funny. It's got a 12V-2 power connector. It has a metal back plate to
help with cooling for the VRAM because it's a switch to a clamshell design. They can't upgrade your existing card. It doesn't work like that. You have to take the GPU off and put it on an entirely new PCB with entirely new uh you know, designs. This is amazing. Also, if I might be so bold as to recommend you do it with two if you have two 4090s. It's a very worthy upgrade. I will be seeing how it stacks up against the uh RTX A6000. Twins. Twins, bastille. This is definitely the most affordable way to get uh 48 gig GPU upgrade if you already have a 4090. Where can you have your GPUs upgraded, too? Uh it's GPU Lab. It's Greg's GPU
Lab. Check it out. They have different upgrade services and some other things they offer. I reached out to get this done incognito. Might have failed on the incognito part, but can't beat the service and the build quality. So, you got 96 gigs of GPUs. You got to put it in a display case, right? This is the Fractal case we reviewed previously. Got it in an AM5 system as well. These are GPUs only going to be connected at X8. It's kind of important because inter-GPU transfers not via NVLink.
It's via PCIE. I don't know what some of you are going to say. The PCIE to PCIE peer-to-peer thing on the 4090, Nvidia turned that off. They did. What you talking about, Con? Pretty sure we're able to do I like the dual slot design because you got ample breathability between our two blower style GPUs. And yeah, they will be a little bit louder, but only under load. No, no, you don't need to adjust your set. It is about 59 to 60 dB from 18 inches away. And that is with the AIDA64 stress test.
Now, when you're doing inferencing or AI workloads, it's much quieter. Around 45 to 48 dB, give or take, when it's under sustained load. The hot spot for inferencing workloads, just, you know, running LM Studio tensor parallelism two, it's something like 60°C, but even with this full load, it's hovering around 81°C. Maximum temperature here is 84°C. So, it's staying under its maximum, even though it's on a new heat sink and card and everything else. The blower motor is very effective at removing heat. And just to be clear, if you're interested in doing a comparison to an NVIDIA reference A6000, the Franken 4090 beats it all to pieces.
Better cooling, better fan profile, better card design from a cooling perspective. Our brain for 64 shows that pretty clearly, I M H O. Okay, so I've got this thing updated to Linux now, and we've got some pretty interesting results. There's a forum thread that has much more detailed info, but our 4090 is uh a fair bit faster than our RTX A6000, at least in this configuration. And this is the ampere generation, not the Ada generation, cuz I mean, I don't even I had some ampere RTX 5880s that I did some horse trading for, and those are great. Uh but the only reason I have these Ada generation cards is because some crazy
audience members sent them to me. Cuz they upgraded to Ada from ampere. It was like, here, have my old ampere cards. And they've been running continuously. You can see projects from uh ubergarm and uh UOS on the forum on huggingface, and that's been a lot of fun. And so, like all this stuff gets used to run pretty much continuously. I'm sort of shocked at how much better that I mean, I get it. I get why people are upgrading 4090s now. So, check out that forum thread. We test small dense models, which is really what I'm after, and the 4090 really shines running small dense models. Not a ton of power consumption, nowhere near our 450 W power limit for the card, and our temperatures stay well under 70° C
uh even at moderate noise levels. Nowhere near the worst noise levels that we uh encountered doing torture testing or whatever you want to call it. Uh noise levels it's about the same as a normal Nvidia design card. I noticed the card that we got back that has my chips on it. There's no Nvidia branding, there's no Nvidia anything, there's nothing. It's just literally take my chips and put them on this other board. And it is sort of fun and delightful in the fact that you can I mean, you can have two of these GPUs for like 2/3 to half the cost of an RTX 6000. And that's one way like an RTX 6000 has 96 gigs of VRAM. Boom, there's 96 gigs of VRAM in this machine. And RTX 6000s are now
selling for over $10,000. If you already have a 4090 and you can upgrade it for like 15, 16, 1700 dollars, that's, you know, a thing. Also, uh you can get water blocks. I am probably going to do another couple of these and probably going to do the water block version of that cuz you know, that's easier to deal with. You can just get a pre-shipped with a nice water block. So, it's going to be an option. Oh, yeah, uh the other thing is like 96 gigs of VRAM doesn't make sense to do that when you can get 128 gigs of VRAM in something like the GMK Tech this is the Evo X3 there's also Evo X2 all kinds of different versions of Strix Halo with 128 gigs of VRAM. Uh yeah, this is LPDDR5. This is much, much faster for
token generation and decode. This is VRAM. This is just regular RAM. This is the bandwidth differences there are significant. But this does cost less than this. Even if this is like on the order of 4 grand now. I don't know. We've got experiments to run. Anyway, the other aspect of this that makes me hopeful for the future is that this is recycling and upgrading enterprise hardware. There've been some folks that have been running enterprise hardware for gaming and everything else and this is recycling the hardware. It requires soldering and a new PCB. That might be what our future is if all of the manufacturing for everything else stops. And don't forget things like the BC250.
This can do 1080p double A titles that approaching 60 FPS now that it's unlocked. It is an uphill battle with the BC250. Um but you can use it as a gaming PC ish. Really on the Linux side more than the Windows side with the unlocks and everything else, but this is another example of recycling hardware in a way that it was never meant to that requires soldering and unlocks and everything else. So, anyway, yeah. So, upgrading your GPU to 48 gigs of memory without having to fly to Taiwan is a thing that I've experienced. So, if you want to do this, uh hang out in the forum. Show pictures of your machine. 46 gigs.
All right. I'm signing out and I'll see you in the forum. Woo, upgrades.
