For PC gaming news and discussion. PCGamingWiki
Rules:
I bought a Lenovo thinkpad x1 carbon from a business closeout dealer or whatever for $150 is the greatest computer purchase I have ever made. It's running Linux now, and I upgraded it with a 2tb ssd, and its tiny and fast, 16gb ram, hd display and fantastic keyboard.
I love the idea of the framework and they seem awesome, but when such a cheap option exists it is hard to justify to insane price tag. Desktop will only be worse I would imagine.
I have a x1 carbon gen 9. Be very careful with the USB-C ports. I leave mine on my desk and my hub kept cutting in and out while using it and its gotten worse over time. I've seen others with similar issues and it seems like the fix is to solder on a new port. Its a work laptop so I just deal with it.
Ah, yes, the one with the checks notes soldered on RAM.
Yes the one with 128GB of ram usable by the graphics card with every other component down the the usb slots modularised and swappable. Its a deliberate trade off that makes it very attractive for people doing ML work, its just with that level of grunt it's also good for games.
Isn't the point of desktops to already be modular?
The funny thing about this new framework is that it isnt
Jup, soldered in RAM. They're less modular than most laptops
Right. Of all the products the company known for repairability and upgradability could have developed, they went into a class of products that didn't need it and removed it.
In their defense though, this was a product class that was missing. If you're an idiot and care at all about AI, this will be an excellent product for that with the 256GBPS RAM bandwidth at half the price of an Apple product.
So maybe they're like Lambo and Porsche who sell giant SUVs to fund the development of their sports cars.
No one should buy this but as long as they keep selling repairable laptops, it's ok I guess.
Yeah finally watched the review. Weird, but also, this is actually a really exciting chip.
If you do any kind of work with ml, this might just about be your only option for breaking 16gb vram, which is the floor for many models. Very few of us ever even get to experiment with something above 40. With these, you could take it to 128, and it seems, at a pretty price competitive standpoint.
I'm looking at it pretty seriously, because I was really excited about these chips, but almost completely disinterested in them as far as a laptop chip. But for a desktop chip? I mean if its at 128gb vram, the fuck else can I ask for?
I'm also super interested in that 2:1. For mobile computing thats pretty much my go-to these days, is a thin and light 2:1 that I just use to get access to where-ever my compute is actually happening.
Maximum allocation of 96GB to VRAM on the 128GB configuration, but your point still stands. This desktop was absolutely designed almost specifically for ML-enthusiasts, and if you wanna run a game on it you can too. Describing it as a "gaming PC" is totally missing the mark.
EDIT: it has been pointed out that the 96GB limit is a Windows limitation, so wouldn't affect any serious ML-enthusiast
The 96GB limit is just for Windows. It can be taken higher on Linux.
96GB on Windows, configurable to more on Linux.
I wouldn't necessarily say it was designed specifically for ML people though the 128GB spec will definitely draw in that crowd, the 32GB model is $1,099 and competes well in the small but very real "Gaming NUC" space that's been dominated by Intel/Nvidia laptop gear in tiny desktop cases. Asus took over the NUC line, and the gaming models are priced way above this without the same ML draw of unified RAM.
Oh, interesting! That's the first I'm hearing of being able to configure more in Linux, seems like anyone taking ML seriously would be using Linux anyway.
Thanks for that correction.
96gb of VRAM? Even most ML professionals have never seen that much vram in their life.
Yeah to me this is like SGI marketing their computers as gaming PCs
If you do any kind of work with ml
I'm looking at it pretty seriously
Im curious what kind of practical application an individual might even have for ML?
That seems like the only legitimate use-case for this device, but also I consider most (not all) ML illegitimate and pointless at the same time.
I consider most (not all) ML illegitimate and pointless at the same time.
Myopia is a curable disease. Why should I bother treating someone so intentionally ignorant with the grace of charity?
Im just going to interpret that unwarranted rude response as not having an answer. I did not ask for charity, I asked a question.
I did not ask for charity
Brother I don't need a Wikipedia link to know what charity is. I'm a native English-speaking adult. That aint it.
I mean apparently you do, considering you don't know what you are asking. You are asking me to make a preposterously generous interpretation of your question, when you yourself can't be bothered to even consider what might be a useful application of ml. Give _ me _ a _ break.
Go sort things on an abacus and get away from anything related to computation you Luddite.
Hey fellow autist, this isn't a good look. Tone it down a bit.
I mean apparently you do, since you don't know what I'm asking.
Yeah I am a huge Framework fan (have one of their laptops and love it) but I at first had a hard time figuring out who this was for. Some other people in this thread have had some good comments about it, but for me personally this is completely useless.
It's for me. I guess since I pre-ordered. Why because I want to mess with llms in it and the 96 gigs of vram seem nice.
As one of the people with a different viewpoint that is a completely fair and sensible take.
Looking at the others in the ultra-SFFPC market segment they're targeting (e.g. Mac Mini, Intel NUC, Nvidia DIGITS) this is a solid first outing.
It's a standard ITX mainboard that happens to have soldered ram. It will fit in any ITX-compatible case and even has dedicated PCI-e slot in case you do use a case with space for a PCI-e device like an SFP+ card.
On the upside, the unified ram means the GPU can use it, and so you could run 70b-size models on it.
On the upside, the unified ram means the GPU can use it, and so you could run 70b-size models on it.
The version with 128GB ram is $1999.
Nvidia's equivalent (DIGITS) is $2,999.
How do you know that it is equivalent? Or what the price is?
Equivalent as in 128gb of unified RAM targeting casual ML workloads. The price should be on any news article about it.
I mean there's more to a computer than how much RAM it has.
I only see 3rd party sources for pricing, nothing directly from Nvidia. And indeed Framework did say in their presentation that there was no price.
I mean there's more to a computer than how much RAM it has.
Ok bro
I only see 3rd party sources for pricing, nothing directly from Nvidia. And indeed Framework did say in their presentation that there was no price.
Terrible mobile website asking to turn off the adblocker making the entire thing unusable. Can't be bothered to go through archive.is to make it readable....