[Tepix]

OK, here's my quick critique of the article (having built a similar AM4-based system in 2023 for 2300€):

1) [I thought] The page is blocking cut & paste. Super annoying!

2) The exact mainboard is not specified exactly. There are 4 different boards called "ASUS ROG Strix X670E Gaming" and some of them only have one PCIe x16 slot. None of them can do PCIe x8 when using two GPUs.

3) The shopping link for the mainboard leads to the "ASUS ROG Strix X670E-E Gaming" model. This model can use the 2nd PCIe 5.0 port at only x4 speeds. The RTX 3090 can only do PCIe 4.0 of course so it will run at PCIe 4.0 x4. If you choose a desktop mainboard for having two GPUs, make sure it can run at PCIe x8 speeds when using both GPU slots! Having NVLink between the GPUs is not a replacement for having a fast connection between the CPU+RAM and the GPU and its VRAM.

4) Despite having a last-modified date of September 22nd, he is using his rig mostly with rather outdated or small LLMs and his benchmarks do not mention their quantization, which makes them useless. Also they seem not to be benchmarks at all, but "estimates". Perhaps the headline should be changed to reflect this?

[jychang]

Yeah, this page seems to be not great for beginners and also useless for people with experience.

A 2x 3090 build is okay for inference, but even with nvlink you're a bit handicapped for training. You're much better off with getting a 4090 48GB from China for $2.5k and just using that. Example: https://www.alibaba.com/trade/search?keywords=4090+48gb&pric...

Also, this phrasing is concerning:

> WARNING - these components don't fit if you try to copy this build. The bottom GPU is resting on the Arctic p12 slim fans at the bottom of the case and pushing up on the GPU. Also the top arctic p14 Max fans don't have mounting points for half of their screw holes, and are in place by being very tightly wedged against the motherboard, case, and PSU. Also, there's probably way too much pressure on the pcie cables coming off the gpus when you close the glass.

[someperson]

What an indictment on NVidia market segmentation that there's an industry doing aftermarket VRAM upgrades on gaming cards due their intentionally hobbled VRAM.

I wish AMD and Intel Arc would step up their game.

[writebetterc]

Intel Arc Pro B60 will come in a 48GB dual-GPU model. So yeah, hardware is gonna be there, and the 24GB model will be $599 from Sparkle. I assume 48GB will be cheaper than a hacked RTX 4090.

Look at this: https://www.maxsun.com/products/intel-arc-pro-b60-dual-48g-t... https://www.sparkle.com.tw/files/20250618145718157.pdf

[magicalhippo]

Keep in mind that the dual-GPU is done via PCIe bifurcation, so that if you use two B60's on a similar motherboard to what's in the article, you'll only see two GPUs, not the full four. Hence just 48GB VRAM not 96GB.

[mertleee]

Yeah, but the B60 is basically half the speed of a 3090... in 2025. I'd rather buy 5yr old nVidia hardware for $100 more on eBay than an intel product with horrendous software support that's half the speed effectively. This build is so cool because the 2x 3090 setup is still maybe the best option 5yrs+ after the GPU was released by nVidia.

[Gracana]

$2.5k is about $1k more than you'd spend on a pair of 3090s, and people I know who've bought blower 4090s say they sound like hair driers.

[dotnet00]

Blowers are loud, but they're easier to pack together, particularly given how most motherboards don't seem to space their two slots sufficiently to accomodate the massive coolers on recent GPUs.

[mertleee]

I can't wait for blower 3090s from China / MSI to get cheap (although I fear this may never happen)

[tacomagick]

Simply replacing the 3090's with 4090's would provide a major performance uplift assuming your model fits. (I have rented both 3090 and 4090 systems online for research, this comment is based on my personal experience, it is well worth the price increase and the hourly rate for the inference speed you get)

[tryauuum]

I am not a lawer, but shouldn't 4090s be worse since they don't have nvlink?

there are patched drivers for enabling p2p but if I remember correctly, they are still slower than having an nvlink

[Aurornis]

Don’t those modified cards require hacked drivers? I would not want my expensive video card to depend on hacked drivers that may or may not continue to be available with new updates.

[nxobject]

Are the Alibaba 4090s modded to reach 48GB VRAM? (I ask only to figure how why they're that cheap...)

[noboostforyou]

Yes, they are modded by replacing the individual VRAM modules. https://www.tomshardware.com/pc-components/gpus/usd142-upgra...

[hengheng]

I've also learned the hard way to Google "AM4 main board tier list" before buying.

Some boards can run a 5950X in name only, while others can comfortably run it close to double its spec power all day. VRMs are a real differentiator for this tier of hardware.

(If anyone can comment on the airflow required for 400-500W Epyc CPUs with the tiny VRM heatsinks that Supermicro uses, I'm all ears.)

[fkyoureadthedoc]

> The page is blocking cut & paste. Super annoying!

I've been running Don't F* With Paste* for years for this

https://chromewebstore.google.com/detail/dont-f-with-paste/n...

[mertleee]

Hmm, I can copy paste just fine from the build page?

[fkyoureadthedoc]

I don't know if the page actually f's with copy/paste or not since I already have the extension. It's usually most useful on forms where they force you to type in stuff.

[jgalt212]

Interesting. I guess our content-based marketing pages need to move to canvas-based rendering. That's probably bum too. Straight to serving up jpgs.

[vidarh]

> Straight to serving up jpgs.

Back in my Amiga-days we had PowerSnap[1] which did the bargain basement version of OCR: Check the font settings of the window you wanted to cut and paste from, and try to match the font to the bitmap, to let you copy and paste from apps that didn't support it, or from UI element you normally couldn't.

These days, just throwing the image at an AI model would be far more resilient...

I think we've gotten to the point where it would be hard to compose an image that humans can read but an AI model can't, and easy to compose an image an AI can read but humans can't, so I suspect the only option for your marketing department will be to try to prompt inject the AI into buying your product.

(Oh, look, I have written nearly this same comment once before, 11 years ago, on HN[2] - I was wrong about how it worked, and Orgre was right, and my follow up reply appears to be closer to what it actually does)

[1] https://aminet.net/package/util/cdity/PowerSnap22a

[2] https://news.ycombinator.com/item?id=7631161

[93po]

thankfully most web browsing will be done by LLMs soon and that won't stop them, good riddance to the mess of a web that google has created

[jgalt212]

dead Internet for realz

[yjftsjthsd-h]

> 3) The shopping link for the mainboard leads to the "ASUS ROG Strix X670E-E Gaming" model. This model can use the 2nd PCIe 5.0 port at only x4 speeds. The RTX 3090 can only do PCIe 4.0 of course so it will run at PCIe 4.0 x4. If you choose a desktop mainboard for having two GPUs, make sure it can run at PCIe x8 speeds when using both GPU slots! Having NVLink between the GPUs is not a replacement for having a fast connection between the CPU+RAM and the GPU and its VRAM.

Forgive a noob question: I thought the connection to the GPU was actually fairly unimportant once the model was loaded, because sending input to the model and getting a response is low bandwidth? So it might matter if you're changing models a lot or doing a model that can work on video, but otherwise I thought it didn't really matter.

[Tepix]

In general, if all you do is inference with a model that’s in VRAM, you’re right. OTOH it’s simply a matter of picking the right mainboard. If you have one of those sweet new MoE models that won‘t completely fit in your VRAM, offloading means you want PCIe bandwidth, because it will be a bottleneck. Also swapping between LLMs will be faster.

[danparsonson]

> None of them can do PCIe x8 when using two GPUs.

Is that important for this workload? I thought most of the effort was spent processing data on the card rather than moving data on or off of it?

[glax]

Sorry for going off topic. But your insight will be helpful on my build

I'm thinking about a low budget system, which will be using

1.X99 D8 MAX LGA2011-3 Motherboard - It has 4 pcie 3.0 x16 slots, dual cpu socket. They are priced around $260 with both the cpu

2. 4X AMD MI50 32G cards - They are old now, but they have 32 gigs of vram and also can be sources at $110 each

The whole setup would not cost more than $1000, is it a right build ? or something more performant can be built within this budget ?

[juliangoldsmith]

I'd use caution with the Mi50s. I bought a 16GB one on eBay a while back and it's been completely unusable.

It seems to be a Radeon VII on an Mi50 board, which should technically work. It immediately hangs the first time an OpenCL kernel is run, and doesn't come back up until I reboot. It's possible my issues are due to Mesa or driver config, but I'd strongly recommend buying one to test before going all in.

There are a lot of cheap SXM2 V100s and adapter boards out now, which should perform very well. The adapters unfortunately weren't available when I bought my hardware, or I would have scooped up several.

[BizarroLand]

I've seen the sxm2 (x2) with pci extension cards out on ebay for like $350.

The 32gb v100s with heatsink are like $600 each, so that would be $1500 or so for a one-off 64gb gpu that is less overall performant than a single 3090.

[OakNinja]

Better to buy one used 3090 than those old cards. Everything is not vram. Or, you can do nothing without vram but you can’t do anything with just vram.

To use the second pair of pcie slots, you _must_ have two cpus installed. Just saying in case someone finds a board with just one cpu socket populated.

[ericdotlee]

Any reason you wouldn't opt for the 4090 or 5090?

[Instantix]

3090 second hand can be found at something like $600.