Test Report — AMD Radeon RX 6700 XT

AI Inference

AI Capability

Pass

Compute

Detected

Display

46.3tok/s

Generation

543.2tok/s

Prompt processing

944ms

First token

qwen3.5-9b-q6-k · Q6_K · ctx 32768 · vulkan · engine b10698

Self-test (QA): Pass

Sustained AI load: 10 min, 53 requests, 9/9 correctness checks passed

VRAM memory test: passed (10 min)

Under AI inference load

71°C

Peak core temp

87°C

Peak junction temp

187W

Peak board power

2607MHz

Peak core clock

1998MHz

Peak memory clock

Good for: Runs a 9B-class model (qwen3.5-9b-q6-k, Q6_K) with a 32K-token context at interactive speed: local chat, coding help and document Q&A.

Stress test

OCCT 3D stress test

OCCT 3D stress test
15 minutes
Duration
15 min 0 s
Errors found
0
Card identity
verified
Peak core temp
75 °C
Peak junction temp
97 °C
Peak memory temp
77 °C

Temperatures and power from OCCT's own sensor log.

Thermals & airflow

Runs warm — and how warm is up to your build.

This is a fully functional card that runs warmer than average. Real-world temperature depends heavily on your case airflow and fan setup.

To show what's possible, we tested with the free Lightweight AMD GPU Fan Control utility engaged — you can see the drop in the chart above. In a well-ventilated case, expect cooler still.

Fan control engaged at 05:53 into the test, when the core averaged above 74 °C.

Before: core 74 °C, hotspot 93 °C, fans 34 %.

With fan control: core 70 °C, hotspot peak 90 °C, fans about 42 %.

Result: held at or below 74 °C.

MetricStockFan Control

Get Lightweight AMD GPU Fan Control

Methodology

How this unit was tested

OCCT 3D stress test (GPU load + error detection)
A 15-minute sustained full-load 3D stress test with automatic error detection.
memtest_vulkan VRAM integrity test
Vulkan-based memory test that checks the GPU's VRAM for integrity errors.
AI inference benchmark
Measures how fast this GPU generates text and processes prompts with a local language model, and runs a short self-test of the answers.
10-minute sustained AI soak with 9 correctness checks
Runs the AI workload continuously to confirm speed and answers stay steady over time.