Is NVIDIA still the default best choice for local LLMs in 2026?

[-]

BoogerheadCult@reddit

Nope, the NVIDIA fanboys are pushing prices to very unreasonable levels, $/VRAM is almost 2x, 3x competitors, I am looking at other alternatives now.

The breaking point for me was seeing how badly used 3090 which sellers didn't even hesitate to disclose that they were used for mining now being sold for $1200. You gotta be kidding me.

[-]

SolemnFuture@reddit

I'm thinking of buying a 7900 xtx or arc pro b70. Dunno which.

[-]

Momsbestboy@reddit

R9700 is in the same price range of the B70, but faster

[-]

New_Comfortable7240@reddit

Amd have better support. I have intel gpu in my laptop and support is lacking

[-]

Ordering 3080 20GB directly from Alibaba costs 500 eur per card when buying a pair,s hipping and tax included. To me, that's undisputed price/performance king. You can save roughly a 100 eur by opting for 2080Ti 22GB instead, and lose significant prompt processing performance due to older tensor cores. Buying those cards makes infinitely more sense than both brand new, and used unmodded options. AMD/Intel can't compete with this price/performance ratio either, let alone on software compatibility.

[-]

Happy_Brilliant7827@reddit

There are some cutting edge tools that only work on mLx macs with unified memory.

Like moe routing on qwen3.6

[-]

RoaRene317@reddit

Inference? No

Training / Finetuning? Yes.

[-]

Kal-LZ@reddit

My Dual R9700 32GB setup is the best investment I've made for just about 2500€ plus VAT

[-]

HaggardSummaries@reddit

Yes. Lot of cope in these comments though.

[-]

usa_reddit@reddit

M series or Nvidia are the two choices

[-]

Vaguswarrior@reddit

I'm using a mixed Nvidia+ AMD Frankensetup. 🤷🏽‍♂️

[-]

DavidBolkonsky@reddit

Same, 5070ti + 9070

[-]

TheHiveFather@reddit

Same, 5090, GB10, Halo Strix.

[-]

Loud-Swim-2932@reddit

NVIDIA often is the simple solution I guess.

[-]

llama-of-death@reddit

At the moment it's great, but I have a feeling much will change before we hit 2027. There are breakthroughs in methods that bypass GPU altogether, and that will shake the Nvidia shareholders big-time.

[-]

totosse17@reddit

Not default, but 90% of the time. I summarized it there: https://llmrequirements.com/state-of-local-ai/

[-]

libregrape@reddit

Great resource. Thank you!

[-]

rwa2@reddit

For experimentation with cutting edge features, yes.

For raw cost per token/s for an established pipeline, usually not.

[-]

totosse17@reddit

If you compare tps/price Nvidia options still pull out ahead. They cost more but they give more tps as well. Only diff is spark vs halo, but there the diff is prompt processing.

[-]

opossum_cz@reddit

I use Halo Strix and I am not sure if anything can compete with it, it is not fast, but it can load large models, price is not that high and doesn't have high power consumption.

If AMD releases something with higher memory bandwidth and 256GB RAM on high frequency, it would be killer.

[-]

Sooperooser@reddit

How does the M4 only use 22w?!

[-]

noctrex@reddit

They where building up the CUDA ecosystem for 2 decades. They are essentially untouchable. For many years to come. Nothing comes even close. You can see this also on the market share between the brands. They are essentially a monopoly now.

That said, I went for a 7900XTX. because 24GB on the cheap

[-]