On Tue, 11/3/2026 10:28 AM, Yamn3 Remailer wrote:
Hi,
Greetings,
I have assembled about 2 PC
for personal use in the past
Now I dont have money
to buy PC parts
But I want to learn
and keep updated
of what is going on in the PC industry
I may try to get custom orders
That customer has to pay
for custom built PC
I am located in India
So I dont know if all advanced
technology and parts is available or not
and I am not connected to the industry
Assembling PC was very easy earlier
Can I also do repariing and fixing of PC s
Currently I have an old Windows 7 PC
that went kaput suddenly
I have to check if it is only the problem with
Power Supply
And a working Windows 10 PC with GTX 1060 graphics card
which I dont think I can upgrade easily , more than 10 GB RAM
Mostly I want to upgrade the graphics card
Because new software use GPU
Regards.
Some software, does run on your GTX 1060. I have a GTX 1050 Ti in
the machine across the way, it has only 4GB of RAM, and using
LMStudio and Linux, I have seen the video card load 2GB of a LLM-AI
model and compute on the card. However, any time an open-weights
model "spills outside of VRAM", it has a high impact on the speed
of the LLM-AI execution. For LLM-AI to work, you must have enough
VRAM to hold the model being used at the moment (MoE, Mixture Of Experts).
If you use two video cards and use "sharding" to split the model
and hold it in the two cards VRAM, this requires the support of
"PCI Express Atomics", which is a set of primitives used for
signaling between video cards, that video card to video card
transfer operations are completed. You cannot plug two
video cards into the x16 and x4 slots on a modern motherboard,
and expect LLM-AI to run properly. Only a motherboard that
takes the x16 on the CPU, splits it and makes x8/x8 mode,
supports PCI Express Atomics.
This makes the cost prohibitive, for any normal home user which
is not a business, to be building PCs for that purpose. For
example, you could buy a ThreadRipper Pro motherboard, an expensive
processor, to gain some PCI Express slots to hold multiple video cards.
The price could run $15,000 to $20,000 and the spending just does not
stop. A compromise is the DGX Spark, which for less money, has
enough memory to load some models. But those aren't all that
fast at LLM-Ai tasks. And they can have performance issues
you cannot measure, in some cases. Things might be slower
than you would like.
*******
If you are setting up a gaming machine, the cheapest card at my
local shop is $1000. All the cheaper cards (like that GTX 1050Ti)
for $200-$300, those are all gone. There are
- no cheap hard drives (price doubled sing January)
- no cheap video cards
- DDR5 is outrageously expensive ($5000 for two 64GB DDR5 sticks)
- motherboards are rising in price (it is claimed PCB substrates
are getting hard to acquire by the motherboard companies, which is
just a lousy excuse). Looking at the $1000 video card I bought,
you can tell from the color of the PCB material, first rate
stuff is not being used, it is reserved for datacenters.
Really, no aspect of computing is affordable.
There will be refurbished machines, but, those are not cheap either.
Naturally, the refurbisher charges... what the market will bear.
The prices were low previously, because the machines were not
all that wonderful.
Roughly speaking, video card prices reflect the "perceived value
of the amount of VRAM on board". That's why a $300 video card
now costs $1000 or $2000. Part of it, is the VRAM it contains.
All "gaming" cards are priced according to the level of
LLM-AI inference they can achieve.
This is why we're not building computers right now.
Financially, not prudent.
Paul
--- PyGate Linux v1.5.19
* Origin: Dragon's Lair, PyGate NNTP<>Fido Gate (3:633/10)