Skip to content
Voltarum Local LLM Hardware Guide

NVIDIA · NVIDIA GPUs · 2020

NVIDIA GeForce RTX 3090

Good budget optionDiscontinued (still sold)

The homelab classic: 24 GB and 936 GB/s, bought used. Two of them make 48 GB.

Memory
24 GB
Bandwidth
936 GB/s
Power
350 W
Largest popular model at 4-bit
Qwen3.6 27B

from $1,340 (typical price)

Researched from manufacturer data and public sources; not tested by us.· Last reviewed Oct 11, 2026

What it can run

Estimate

Popular models at Q4_K_M with a 32K-token context. Memory needed from real model files; speed is the theoretical ceiling.

Change model size or context

Model weightsContext cache (KV)RuntimeUsable limit (colour = verdict)

ModelVerdictMemorySpeed ceiling
Qwen3 1.7BRuns
6.4 GB / 24 GB · with 24 GB
≤ 843 tok/s
Qwen3.5 9BRuns
8.3 GB / 24 GB · with 24 GB
≤ 165 tok/s
Gemma 4 12BRuns
9.5 GB / 24 GB · with 24 GB
≤ 131 tok/s
gpt-oss-20bRuns
14 GB / 24 GB · with 24 GB
≤ 468 tok/s
Qwen3.6 27BRuns
20 GB / 24 GB · with 24 GB
≤ 56 tok/s
Qwen3.6 35B A3BPartly in system RAM
24 GB / 24 GB · with 24 GB
—
Llama 3.3 70BNot enough memory
55 GB / 24 GB · with 24 GB
—
gpt-oss-120bPartly in system RAM
65 GB / 24 GB · with 24 GB
—
Qwen3.5 122B A10BPartly in system RAM
79 GB / 24 GB · with 24 GB
—
Qwen3.5 397B A17BPartly in system RAM
247 GB / 24 GB · with 24 GB
—

Theoretical maximum for generating text: memory bandwidth divided by the bytes read per token. Real runs usually reach about half to three quarters of it.

Where to buy

Prices are editorial references (launch price and a typical street price on the date shown), not live prices. In 2026 the memory shortage has pushed many prices well above launch.

Prices are shown only if updated in the last 24 hours and may change; the retailer’s price at checkout applies. We may earn a commission from these links at no cost to you. Affiliate disclosure

Overview

Editorial opinion

Best for: 24 GB on a budget, and 48 GB builds with two cards for 70B models.

Editorial opinionWould we buy it today?

Yes, from a seller with a warranty. Still the best value per GB of fast GPU memory.

Pros and cons

Editorial opinion

Pros

  • 24 GB with CUDA for far less than new cards
  • Nearly 1 TB/s of bandwidth
  • Supports NVLink between two cards

Cons

  • Used: no or short warranty
  • 350 W and hot memory chips
  • Used prices have risen a lot in 2026

Specifications

Factual data
Manufacturer
NVIDIA
Chip
GA102
Architecture
Ampere
Type
Graphics card
Form factor
PCIe card
Released
2020
Availability
Discontinued (still sold)
Memory options
24 GB
Memory type
GDDR6X
Memory bus
384-bit
Memory bandwidth
936 GB/s
Usable for models (default)
24 GB / 24 GB
Board / max power
350 W
Interface
PCIe 4.0 x16
Power connector
2× 8-pin (varies by card)
Compute stack
CUDA
Models
Any model that fits
Several units for one model
Yes

Source: Manufacturer specifications

Software support

Editorial opinion

How well the popular local-AI tools run on it today (editorial assessment).

  • llama.cppExcellent
  • OllamaExcellent
  • LM StudioExcellent
  • vLLMExcellent
  • MLXNot supported

Measured speeds

Factual data

Measured results with their source. We never invent speeds.

No measured results yet. The table above shows the theoretical ceiling instead.

Alternatives

AMD GPUs

AMD Radeon AI PRO R9700

Memory
32 GB
Bandwidth
645 GB/s

Largest popular model at 4-bit: Qwen3.6 35B A3B

32 GB for about a third of an RTX 5090: the cheapest new 32 GB card.

from $1,550 (typical price)Recommended

AMD GPUs

AMD Radeon RX 7900 XTX

Memory
24 GB
Bandwidth
960 GB/s

Largest popular model at 4-bit: Qwen3.6 27B

24 GB and 960 GB/s: RTX 3090-class memory, bought new.

from $1,358 (typical price)Recommended

Intel GPUs

Intel Arc Pro B70

Memory
32 GB
Bandwidth
608 GB/s

Largest popular model at 4-bit: Qwen3.6 35B A3B

Intel's big Battlemage: 32 GB and 608 GB/s for less than AMD's R9700.

from $1,268 (typical price)Recommended

Related guides