Skip to content
Voltarum Local LLM Hardware Guide

AMD · AMD GPUs · 2025

AMD Radeon RX 9070 XT

AMD Radeon RX 9070 XT

Photo: Chaddy · CC BY-SA 4.0

Recommended

AMD's best gaming card is a solid 16 GB AI card too, with ROCm and Vulkan.

Memory
16 GB
Bandwidth
645 GB/s
Power
304 W
Largest popular model at 4-bit
gpt-oss-20b

from $759 (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 / 16 GB · with 16 GB
≤ 581 tok/s
Qwen3.5 9BRuns
8.3 GB / 16 GB · with 16 GB
≤ 113 tok/s
Gemma 4 12BRuns
9.5 GB / 16 GB · with 16 GB
≤ 91 tok/s
gpt-oss-20bRuns
14 GB / 16 GB · with 16 GB
≤ 322 tok/s
Qwen3.6 27BNot enough memory
20 GB / 16 GB · with 16 GB
—
Qwen3.6 35B A3BPartly in system RAM
24 GB / 16 GB · with 16 GB
—
Llama 3.3 70BNot enough memory
55 GB / 16 GB · with 16 GB
—
gpt-oss-120bPartly in system RAM
65 GB / 16 GB · with 16 GB
—
Qwen3.5 122B A10BPartly in system RAM
79 GB / 16 GB · with 16 GB
—
Qwen3.5 397B A17BPartly in system RAM
247 GB / 16 GB · with 16 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: Gamers on Linux who also run 8–14B models.

Editorial opinionWould we buy it today?

Fine if you prefer AMD; for AI alone the 32 GB R9700 uses the same chip.

Pros and cons

Editorial opinion

Pros

  • Good bandwidth for the price
  • Works with llama.cpp, Ollama and LM Studio

Cons

  • 16 GB
  • Fewer tools than CUDA; vLLM support is still maturing

Specifications

Factual data
Manufacturer
AMD
Chip
Navi 48
Architecture
RDNA 4
Type
Graphics card
Form factor
PCIe card
Released
2025
Availability
On sale
Memory options
16 GB
Memory type
GDDR6
Memory bus
256-bit
Memory bandwidth
645 GB/s
Usable for models (default)
16 GB / 16 GB
Board / max power
304 W
Interface
PCIe 5.0 x16
Power connector
2× 8-pin (varies by card)
Compute stack
ROCm / Vulkan
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.cppGood
  • OllamaGood
  • LM StudioGood
  • vLLMWorks with effort
  • 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

NVIDIA GPUs

NVIDIA GeForce RTX 5070 Ti

Memory
16 GB
Bandwidth
896 GB/s

Largest popular model at 4-bit: gpt-oss-20b

Nearly RTX 5080 bandwidth for less. Still 16 GB, so speed rather than size.

from $1,190 (typical price)Recommended

NVIDIA GPUs

NVIDIA GeForce RTX 5060 Ti 16 GB

Memory
16 GB
Bandwidth
448 GB/s

Largest popular model at 4-bit: gpt-oss-20b

The cheapest new NVIDIA card with 16 GB. Low power and CUDA make it the easy starter card.

from $789 (typical price)Good budget option

Intel GPUs

Intel Arc Pro B60

Memory
24 GB
Bandwidth
456 GB/s

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

The cheapest new 24 GB card, if you can live with Intel's software.

from $599 (typical price)For advanced users