Custom Low-Profile RX 9060 XT 16GB Review: Gaming, AI and Noise vs. Arc Pro B50

1 / 5 · The custom low-profile RX 9060 XT blower card in the test build.

I finally have numbers for that half-height RX 9060 XT 16GB I posted. I also tested an Intel Arc Pro B50 in the same SFF build to see how they compare.

I bought the RX from GGcİMIDA, an independent modder in China. Credit for the cooler design and machining goes to the modder. I handled the build and testing, including finding out how loud the blower gets.

The donor is a VASTARMOR AH-9060XT16DCBR3N. It already has a half-height PCB under the full-height stock cooler. The card retains the donor model’s power settings with the custom cooler fitted. Driver-reported board power got close to 180W in some tests. The official LP card is still hard to get, and I have a soft spot for blowers, so here we are. I haven't tested the official LP version in this system.

The RX was faster in Cyberpunk, FFXIV, all five 3DMark tests, and LLM token generation over Vulkan. It also ran hot.

Test setup

CPU: AMD Ryzen 5 9600X
Motherboard: GIGABYTE B650I AORUS ULTRA
RAM: ADATA XPG LANCER BLADE DDR5-6000 CL48, 32GB (2 × 16GB)
PSU: Custom fanless Enhance 7660L with a copper enclosure
Case: DENG 1U Mini-ITX case
CPU cooler: SilverStone NT06-PRO V2, with a Cooler Master Mobius 120 Slim underneath, blowing upward

ReBAR was enabled. The RX used Adrenalin 26.8.1; the B50 used the standard Arc 32.0.101.8992 driver. I haven't tested the Pro driver here.

Gaming and 3DMark results are averages of three valid runs, with default clocks, default power settings, and automatic fan control. The 140W tuning I mentioned in the previous thread wasn't used for these results.

1440p gaming

In Cyberpunk 2077's built-in benchmark, the RX averaged 67.66 FPS against 35.95 FPS on the B50. FFXIV Dawntrail came out at 91.89 vs. 59.74 FPS. A pretty noticeable gap in both.

Both ran at 2560 × 1440. Cyberpunk used Custom settings based on the Ultra preset, with upscaling, frame generation, ray tracing, and path tracing off. FFXIV used High (Desktop), FSR at 100%, and dynamic resolution off.

Cyberpunk 2077 and FFXIV Dawntrail: averages of three valid runs at 1440p.

I ran Black Myth: Wukong too: 36.67 FPS on the RX and 18.00 FPS on the B50, averaged over three runs. The saved render-scale settings differ between cards, though, and I couldn't confirm they matched during the actual runs. I've kept these results separate.

Black Myth: Wukong, reported separately because matching render-scale settings could not be confirmed.

The RX also leads in all five 3DMark tests below. Time Spy and Time Spy Extreme use graphics scores; the other three use their respective benchmark scores.

3DMark

Time Spy and Time Spy Extreme graphics scores.

Steel Nomad, Port Royal, and Speed Way benchmark scores.

Local LLMs

For this part I used LM Studio 0.4.23+1 with llama.cpp Vulkan 2.32.0: GPT-OSS 20B, Qwen3 4B Thinking/Instruct, and Gemma 4 12B IT. The tasks were Pizza, Shopping Cart, E-commerce, and Itinerary, plus an extra GPT-OSS Shopping Cart test.

Each workload got a warm-up followed by three measured runs. Settings were an 8192-token context, 1024-token output cap, temperature 0, seed 42, and maximum GPU offload.

Across these five workloads, the RX generated tokens 1.82–1.99× as fast as the B50 with this Vulkan setup. Gemma, for example, averaged 35.46 vs. 19.46 tok/s.

LLM generation speed with llama.cpp Vulkan: GPT-OSS and Qwen3 workloads.

Gemma and the extra GPT-OSS Shopping Cart workload. The extra test produced different output lengths: RX 865 tokens, B50 1024 tokens.

It used more power doing it: average driver-reported board power during the AI requests was about 130–158W for the RX and 67–69W for the B50.

Heat, noise, and the takeoff button

The RX hit 110–112°C hotspot in some of the gaming and 3DMark logs, with memory peaking at 106°C. Those temperatures at default settings are my main concern.

The card has fan stop at idle, and the fan curve is adjustable in AMD Software. Under load with automatic fan control, it runs around 4,000 RPM. From my usual seat, I barely notice it.

Then I manually set it to 100%, roughly 9,000 RPM. That was me pressing the takeoff button. I measured about 70.6 dBA for the whole system during the full-speed Time Spy segment. Here’s a clip with the fan deliberately maxed out:

Time Spy with the fan manually set to 100%, roughly 9,000 RPM. Twenty seconds of original audio, with no volume adjustment.

Recorded peak temperatures. The room was held at 28°C for both cards.

Recorded peak fan speeds with automatic fan control; manual full-speed runs are listed separately.

Whole-system noise readings at the recorded fan speeds and loads. Shared background: 49.8 dBA.

I used a GM1351 sound level meter at 15 cm. The readings include the whole PC and background noise. I don't have a usable dBA measurement for the RX at roughly 4,000 RPM under load with automatic fan control, so I can't say which card is quieter at default settings.

I still really like this thing. Half-height, 16GB, fan stop at idle, and the performance is there. So far, the heat may be my only real gripe. There's still some tinkering to do with the power settings and fan curve.

The detailed charts, test conditions, and measurement videos are now up on LabData, my hardware test database, so you can check the results against the original recordings and logs.

The build in the photos will get its own post too. The magnetic wooden cover comes off for quick PCIe card swaps without taking the whole case apart. For now, that’s the GPU testing. I’ll save the wooden case for its own post.