Random Article of the Day...

[revelator]

aye many comapanies who are reliant on memory have taken a huge hit. but not only them it is the entire home computing scene that is getting hit, everything from coolant paste to network cards have seen a huge drop due to most buyers going for the new stuff which is now so expensive that they would rather wait it out (if it ever comes back since many of the AI investors are now trying to force a subscription model through).

you be amazed at how far an old setup can go if you play your cards right . bottlenecks are not to bad with haswell if you stick to 200x series and with the skylake you can even go for the 300x models. all also depends on the resolution you play in. if you as you say go for 2K then the skylake will have no problems at all, 1080p will have some bottlenecking dependant on the card.

[revelator]

funny you mentioned the 3090, it is indeed the goto card for local users of LLM's due to its massive compute power (usually run in 2x nvlink setups). also one of the reasons it is now insanely expensive used.

yeah the AI craze has gone completely of the rocker.  altman is also between a rock and a hard place since he partnered up with microsoft to power the built in AI in windows. his company is literally several billions in debt allready so if he cant score a win for microsoft he will be toast.

also two 3090 does not even get close to what a 5090 costs here 79K!!! thats the price of a small car, a new one at that sigh.

[revelator]

a used 3090 here is around 8 to 9k in danish currency.

in comparison the 3080 costs around 3k used the ti model around 4k.

downside of the 3080 is the lack of vram the card itself is plenty powerfull.

[revelator]

im using a 4790k devils canyon haswell to drive my RX5700XT and it does ok though im only playing games without raytracing on it (does not support hardware raytracing at all so a bit moot). hzd runs flawlessly in 4k on this setup and forbidden west also runs good if you lower settings a bit. the cpu is a bit of a bottleneck at lower resolutions but not as bad as i feared. i would not call it a true retro setup since it runs win10 to me it is more of a media PC with some gaming ability . my real retro machine runs win 7 32 and 64 bit uses 2 geforce 560 ti in SLI has 16 gb ram and is driven by an old broadwell 6800k 6 core cpu. the 560 ti is from around the same time as the first core i 7 nehalem models so the broadwell means im in no way cpu bottlenecked. runs the old crysis 1 2 and 3 on max settings. my developer machine is driven by a 6950x broadwell 10 core and 32 gb ram. it has 1 nvme and one SSD as well as two large mecahnical harddrives i use for source code and my msys2 driven windows clang compilers (to big to fit on the main drive). it uses my rtx 3070 (sold the 2080 ti) and is pretty powerfull despite only having pcie 3 and no rebar (plays alan wake 2 with playable fps "45 -> 50" if you dont go nuts with the ray/path tracing effects). the cpu is only slightly slower than the fastest skylake.

my main gamer recently got an RX 9070 XT as an upgrade for the 2080 ti (doubles the speed) but it is seriously hampered by my ryzen 3700x so im looking to upgrade the cpu with either a 5900x or 5700x3d (5800x3d if i can get one for a good price). my ram are also not the best for the setup as i can only get them to 3200 mhz in the xmp profile (3000 mhz standard). but with dram prices atm. that will have to wait unless i can find a used set of 3600 mhz blocks (2x16 or 4x8 for 32 mb). with that in place my gamer should be ready for many more years of service.

and i still got 3 more mainboards left  one micro atx with 16 gb ddr 3 and a haswell core i5 (xp compatible yes i tested it). a gigabyte board with a core2 quad 6600 and 8gb ddr2 ram. and an asrock board which is a bit of a curiosity. can be used with either or both a pcie and/or agp gfx card. cpu is a pentium but the board can take a core 2 as well (not a quad).

this board can actually run win9x (depends on the gfx card if you can find a driver though).

im giving these away for tinkerers, they all work .

[taffernicus]

On 9/29/2026 at 6:18 PM, revelator said:

funny you mentioned the 3090, it is indeed the goto card for local users of LLM's due to its massive compute power (usually run in 2x nvlink setups). also one of the reasons it is now insanely expensive used.

yeah the AI craze has gone completely of the rocker.  altman is also between a rock and a hard place since he partnered up with microsoft to power the built in AI in windows. his company is literally several billions in debt allready so if he cant score a win for microsoft he will be toast.

also two 3090 does not even get close to what a 5090 costs here 79K!!! thats the price of a small car, a new one at that sigh.

it has tons of VRAM as well 

Unfortunately, as far as I remember, you can't use an NVSwitch on the 3090.

But if you run more than two 3090,  looks like you'll need 100-lane PCIe switches.

[taffernicus]

pcie switches reminds me of this 

For the on-premises setup  we won’t be using this consultant’s personal PC, the plan is to use a Mac Studio M5 for the company’s LLM needs.

At least it’s not for generating AI slop photo , art or video , it’s more for large-scale chatbots, inventory management, and information extraction from customer orders.

 

Meanwhile, I’m still doing everything manually (reports and presentations), handling emails (without an AI summarizer), taking meeting notes without AI tools, making form without AI, and writing automation code,n8n and PowerShell scripts (still using Stack Overflow instead of AI) 

lol 

Edited September 30 by taffernicus

[taffernicus]

 

NVLINK... I still remember the buzz around InfiniBand if I’m not mistaken it was after NVIDIA acquired Mellanox

the eyword is "interconnected" and "chained together"

it used to be the go-to solution for something like Oracle Exadata. It used to be a hot topic in the field of high-performance computing and among those involved in MPI programming.

 

But what I’m curious about with ultra-fast networks like that is the congestion handling, packet fragmentation, OS polling settings, TCP offload features, the MTU consensus used (4092 or higher), whether a TCP window algorithm and window scaling are needed since it already has its own flow control, whether jumbo frames should be set to 9000 or not, and what the sizes of the receive and send buffer memories are, etc. (infiniband underlying networking might be totaly different from regular ethernet and alike,i am just recalling what i remember from Tanenbaum’s networking book, Richard W. Stevens’s book, and CCNA)

QSFP28 may be a totally different beast and not the same as infiniband

Seeing technology like this reminds me of the Transputer again , it was quite influential , unorthodox and has a legacy of its own.

Edited September 30 by taffernicus

[revelator]

for gaming nvlink newer really took off, SLI was allready mostly dead by the time it came out but for compute it was a tool of wonder. the old titanX cards had omni directional nvlink (means you could daisychain as many cards as your board supported) cards that came out after the titanX only had mono directional so you were stuck with max 2 cards (guess ngreedia did not want to shoot itself in the foot allowing home users to make AI farms so they split this off for the dedicated AI monster cards -> no chance in hell for a normal person to afford those). the 3090 only supports mono directional but two were still fast enough to get some fun out of toying with a local LLM. and yeah it has plenty of vram  for two cards you are looking at 48 gb.

if you can find a game that supports nvlink the cards will show you the muscle, but there arent many of those about so its a bit of a waste unless you only play those. so for gaming id say go for something else.

 

[revelator]

im probably the wrong person to ask since i dont dabble in AI. only thing i know about it comes from my friend who makes poetry using LLM's. and he uses the chinese model . atleast it gives him some fun as he can be creative without having special skills in that regard.

from what i seen of his creations it works well enough despite him only having an rtx 3070 to run it on.slow as all hell but it does spit out useable stuff (eventually).

[revelator]

by the great cthulhu  christians have invaded the blood of the dawnwalker forums moaning about them being put in a bad light by a game... looking back chritianity was one of the most bloody in human history and they bitch about the game allowing for people to fight back i mean wtf... yes lets have the dark ages back as if we truely need more to bog us down.

i think its time we neutered they're access to the internet .

[JackFarmer]

11 hours ago, revelator said:

by the great cthulhu  christians have invaded the blood of the dawnwalker forums moaning about them being put in a bad light by a game... looking back chritianity was one of the most bloody in human history and they bitch about the game allowing for people to fight back i mean wtf... yes lets have the dark ages back as if we truely need more to bog us down.

i think its time we neutered they're access to the internet .

Got a link?

Edit: NM, found it and I get what you mean, however, best approach would be to do like in the reply below:

Translation: ignore

Edited Friday at 03:06 PM by JackFarmer

[revelator]

unfortunatly those kinda trolls just tend to double down on the crap .

bit like the MAGA cults allthough i think they spelled that one wrong it's MAGGIA.

[JackFarmer]

https://www.ttlg.com/forums/showthread.php?t=153454&p=2543356&viewfull=1#post2543356

Great reaction from heywood.