Custom half-height blower
RX 9060 XT 16GB
Modified VASTARMOR AH-9060XT16DCBR3N
Adrenalin 26.8.1
Same-PC comparison
Intel Arc Pro B50
16GB · Standard Arc driver in this test
Arc 32.0.101.8992
About the custom RX
The RX donor is a VASTARMOR AH-9060XT16DCBR3N with a half-height PCB and a full-height stock cooler. An independent modder in China designed and built the blower cooler; TK Lab assembled and tested the PC. The official LP model was not tested.
TEST PLATFORM
Shared test platform
Same CPU, memory, power supply and case, with ReBAR enabled.
- CPU
- AMD Ryzen 5 9600X
- Motherboard
- GIGABYTE B650I AORUS ULTRA
- Memory
- ADATA XPG LANCER BLADE DDR5-6000 CL48 · 32GB（16GB × 2）
- Case
- DENG 1U Mini-ITX case with aluminum construction and wood panels
- Power supply
- Enhance 7660L custom fanless version · Copper enclosure
- CPU cooling
- SilverStone NT06-PRO V2with a Cooler Master Mobius 120 Slim underneath · blowing upward
TEST PROTOCOL
How the comparison works
- 01
Same PC, swap the GPU
Keep the platform fixed; use the recorded driver for each card.
- 02
Fix each workload’s settings
Games and 3DMark use default clocks/power and automatic fans.
- 03
Keep independent runs
Use three valid formal game/AI runs; keep warmup separate.
- 04
Compare within conditions
Separate manual full-fan tests; mark setting differences and evidence gaps.
EXPLORE THE TESTS
Four parts of the comparison
01
Games & 3DMark
1440pGame resolution
Cyberpunk 2077, FFXIV Dawntrail and five 3DMark tests.
- Game settings
- Upscaling/frame generation specified by test
- Statistic
- Mean of three valid formal runs
Wukong has different saved render scales; supplementary only.
Full settings and method
Default clocks and power settings with automatic fan control; results are arithmetic means of three valid formal runs. The earlier 140W adjustment was not used. Manual full-fan tests are separate from the default group.
Cyberpunk 2077: 2560 × 1440, Custom based on Ultra; upscaling, frame generation, ray tracing and path tracing disabled.
FFXIV Dawntrail: 2560 × 1440, High (Desktop), FSR 100%, dynamic resolution disabled.
Time Spy and Time Spy Extreme use Graphics score; Port Royal, Speed Way and Steel Nomad use their respective benchmark scores.
02
Local AI inference
Vulkanllama.cpp backend
GPT-OSS 20B、Qwen3 4B、Gemma 4 12B IT。
- Context
- 8192
- Maximum output
- 1024 tokens
RX prefill evidence is incomplete; extra cart output lengths differ.
Full settings and method
LM Studio 0.4.23+1 and llama.cpp Vulkan 2.32.0. Models include GPT-OSS 20B, Qwen3 4B Thinking/Instruct and Gemma 4 12B IT. Workloads cover Pizza, Shopping Cart, E-commerce and Itinerary, plus an additional GPT-OSS Shopping Cart case.
Each case has a warmup followed by three valid formal runs. Context 8192, maximum output 1024 tokens, temperature 0, seed 42, maximum GPU offload. Power figures are driver-reported board readings.
03
Temperature & power
Time seriesSensor readings
Compare GPU temperature, hotspot, power and fan speed over time.
- Power source
- Driver-reported board readings
- Statistics window
- Selected recording and time window
Some formal RX runs reached 110–112°C hotspot and up to 106°C memory.
Open load curves →04
Acoustics & original sound
15 cmGM1351 measurement distance
Seven measurement clips, each tied to its fan speed and load.
- Measured sound
- Whole PC, including background
- Excerpt length
- 20–30 seconds each, original audio
RX default-load noise is missing; no default-noise ranking.
Full settings and method
GM1351 at 15cm, measuring the whole PC including background sound. Values are arithmetic means of per-second LCD readings within each measurement window, not isolated-card sound power measurements.
Public videos are 20–30 second excerpts. Source excerpt and measurement windows are labeled separately. Audio has no gain adjustment, denoising or re-encoding. Private screen content is obscured while the sound meter and original audio remain visible and audible.
READ WITH CONTEXT
Where the comparison applies
SupplementaryBlack Myth: WukongSaved render scales differ between cards.+
Black Myth: Wukong has different saved render scales between cards. It remains a supplementary record, not a settings-matched performance comparison.
Different conditionsGPT-OSS Shopping CartRX 865 tokens; B50 reached the 1024-token cap.+
The additional GPT-OSS Shopping Cart case has different output lengths: RX 865 tokens, B50 at the 1024-token cap. RX prefill evidence is incomplete and does not support a reliable ranking.
Not measuredDefault-load noiseNo usable RX reading at about 4,000 RPM; no ranking.+
No usable RX noise measurement is available for default load at approximately 4,000 RPM, so no default-noise ranking is made. Some formal RX tests reached 110–112°C hotspot and up to 106°C memory temperature.
This overview records the test platform for this batch. Per-run conditions and evidence grades remain authoritative. Unverified ambient temperature and per-run panel state are not filled in or inferred from photos.