Hardware favorites
The ranking is the community's — hearts are opinions, speeds are measurements from real rigs. Sort by either and make up your own mind.
Tier opinions are the community's, not gospel. Argue your corner on the wall.
Join the wall →SRuns the big stuff
RTX 5090 32GB
32GB1792 GB/s · 575W · $2400
not measured yet
The enthusiast sinkhole that finally fits 70B-class Q3/Q4 and makes 8B scream at 150+ tok/s.
RTX 4090 24GB
24GB1008 GB/s · 450W · $2000
not measured yet
Peak consumer bandwidth per dollar in 24 GB — the card that made 32B Q4 feel effortless.
RTX 5080 16GB
16GB960 GB/s · 360W · $1300
not measured yet
Blackwell's mid-range: 960 GB/s of GDDR7 in a 16 GB frame for very fast small-model decoding.
RX 7900 XTX 24GB
24GB960 GB/s · 355W · $950
not measured yet
AMD's wide-bandwidth 24 GB: near-3090 speed for the Vulkan faithful, minus the CUDA tax.
RTX 3090 24GB
24GB936 GB/s · 350W · $950
not measured yet
The people's 24 GB card: the single most-run GPU for 32B Q4 in the entire homelab community.
RTX 5070 Ti 16GB
16GB896 GB/s · 300W · $900
not measured yet
The value Blackwell: 16 GB at a hair under a terabyte-per-second keeps 32B offloads peppy.
819 GB/s · 160W · $5999
not measured yet
The homelab endgame: 192 GB at 819 GB/s lets multi-hundred-billion MoE models feel brisk.
800 GB/s · 150W · $4999
not measured yet
Last-gen flag bearer — a 192 GB workhorse for big unified-memory inference at a more negotiable price.
AThe sweet spot
NVIDIA A6000 48GB
48GB768 GB/s · 300W · $4000
not measured yet
The pro-buyer's 48 GB: room for 72B Q4 with zero offload drama and ECC on the side.
RTX 3080 10GB
10GB760 GB/s · 320W · $420
not measured yet
Fast, hot, and 10 GB — a sprightly racehorse for 7–8B models when the price is right.
RTX 2080 Ti 11GB
11GB616 GB/s · 250W · $280
not measured yet
Aged like fine wine — old stock sells for peanuts while still hitting 60+ tok/s on small Q4s.
MacBook Pro M4 Max 128GB
128GB546 GB/s · 100W · $4999
not measured yet
An unapologetic desktop in a laptop — 128 GB at 546 GB/s runs 70B Q4 at ~20 tok/s.
RX 6800 16GB
16GB512 GB/s · 250W · $380
not measured yet
A well-priced 16 GB RDNA2 card that quietly handles 8B Q4 and light 24B offloads.
Arc B580 12GB
12GB456 GB/s · 190W · $249
not measured yet
Intel's olive branch to the local-LLM crowd: 12 GB at 456 GB/s for a very honest $249.
400 GB/s · 90W
not measured yet
The 'good enough Mac' — 96 GB of unified memory that still chews through 32B Q8 and 70B Q3.
BCapable daily drivers
RTX 3060 12GB
12GB360 GB/s · 170W · $220
not measured yet
The budget hero: 12 GB and solid CUDA support make it everyone's first local-LLM card.
Tesla P40 24GB
24GB346 GB/s · 250W · $150
not measured yet
The thrift-shop datacenter relic — noisy and fanless-by-default, but 24 GB for the price of a steak.
RTX 4060 Ti 16GB
16GB288 GB/s · 165W · $450
not measured yet
The quiet couch-dweller: 16 GB in a 165 W card is enough for Mistral Small Q4 beside the TV.
RX 7600 XT 16GB
16GB288 GB/s · 150W · $330
not measured yet
A cool, affordable 16 GB option for modest local models without the high-watt personality.
273 GB/s · 60W · $2399
not measured yet
The serious laptop pick: 48 GB unified memory makes 30B-class MoE models genuinely usable on battery.
DGX Spark 128GB
128GB273 GB/s · 100W · $3999
60 tok/s · 1 rig
NVIDIA's desk-sized mini-supercomputer — 128 GB of Grace-Blackwell memory for big MoE models at 100 W.
RTX 4060 8GB
8GB272 GB/s · 115W · $300
not measured yet
A dainty, cool-tempered card for 4–7B Q4 models — the power bill stays friendly all year.
Ryzen AI Max+ 395 128GB
128GB256 GB/s · 120W · $2500
not measured yet
AMD's 128 GB APU monster — desktop-class unified memory for big models at a fractal power draw.
CStarter territory
Mac mini M4 16GB
16GB120 GB/s · 65W · $599
not measured yet
The gateway drug: a $599 16 GB mini that quietly runs 8B Q4 models at chat speed on 65W.
MacBook Pro M4 24GB
24GB120 GB/s · 40W · $1699
not measured yet
Same M4 brain, more unified memory — room for 14B-class Q4 models plus a browser full of tabs.