So seen someone talking about a PC they want to build, currently working with this for myself this summer and want to make sure it is future proof with good upgrade paths as well as good enough for when I finish uni (going into IT, most likely machine intelligence). Not exactly good with building PC's only ever had prebuilts.
0
PC Build
posted in
Off Topic
Comments:
Threaded
Linear
if you are planning to go into MI then i would suggest you to get a nvidia card
just for cuda,cudnn and other "native" shit
this thread might help - https://www.reddit.com/r/Amd/comments/10673xf/hows_amd_gpus_faring_with_machine_learningdata/
the 4060 is a borderline scam from a gaming perspective, hence the reviews
but its not bad to the point where you should never get it; you have one of the rare use case where nvidia is simply superior
I only dabbled in ML for a paper in regards to QC so I cant exactly provide any in depth insight.
The one thing I can confirm is that lot of super new implementations are made for Nvidia hardware. However at least for my use cases the 3060 (12gb) already had severe memory limitations, to the point where I had to use the universities Server.
If you go into MI I would assume your university offers a similar service. The main question would be how accessible they are for you (waitlists etc).
ML for next year in the course shouldn't be that intensive so I should be able to get away with 8GB card and if I need to, I have a friend that has a 4080 ti that would be willing to help me out just need to think about the year after that so going to go straight into upgrading at the end of next year to a 4060ti 16GB
uni services arent that bad, cloud based but they do have a tendancy to go down
i would prefer my 3060 to the 4060 just for the vram size(im gonna upgrade it soon tho so i would rather wait than shoot the gun on 4060)
if you are going into deep learning then a larger vram is always beneficial over the boost in gpu core performance
see if you can get a 4060ti 16gigs if youre going to train models on it
(what card do you currently have?)
yeah im looking more and more at the part list Ultia sent in (https://uk.pcpartpicker.com/user/WoozyKoala/saved/#view=7jqM3C) and quite liking it, 4060ti 8GB should be easy enough to sell on second hand as well
I went slightly over budget by £10 but pulled off 3 different builds. You can ignore the third option if you intend to use a cloud service for working on large models.
-
A Ryzen 5 7600 paired with a 4060ti. You get robust CPU Cooling with a Phantom Spirit Cooler and 32GB of fast DDR5 RAM on an AM5 Motherboard. This gives you future upgradability and a system tuned for 1080p high-refresh gaming (540Hz) all in an mATX package:
-
A Ryzen 5 5600 paired with a 4070. You get a decent AM4 Motherboard with only 16GB of DDR4 RAM in the same case with the same PSU but with a 4070. You'd have to use the stock cooler unless you want to invest another £20. You sacrifice long-term future upgradability (up until a 5800x3D) for a system tuned for 1440p single-player gaming but perfectly capable with esports titles up to 240Hz:
-
A Core 5 12600K paired with a 4060ti 16GB. This is a productivity and machine learning build first. Extra e-cores for multithreaded tasks and 16GB of VRAM for larger models and data. Go for this option if, and only if, your models/data will exceed 12GB of VRAM whether that's 3D rendering/animation or model training like with SDXL. The downside is that its compute is choked by low memory bandwidth and this card's value sucks for gaming:
I like the first one but the only problem is that the 50 series is coming out so soon after buying this and dont wanna feel like im getting ripped off because of the price dropping lol
definitely going to look into the first one tho
edit: most likely going to actually take the first one and try it, thanks man really appreciate taking the time to make those for me
Discord