XDA XDB’s $2,304 Local AI PC Build: RX 9070 XT, VRAM and Windows Compatibility

What XDB is and how the test worked​

According to its official page, XDB plans new PC builds and upgrades. It checks whether parts fit together, whether the power supply can handle them, and whether the budget goes to the parts that matter most. It also works out which parts you can keep when upgrading. Prices are for US stores. Build types on offer are PC, gaming PC, creative PC and NAS. The budget is optional, and with no budget, parts are chosen on the request alone.

Ty Sherback, the XDA writer behind the test, prompted XDB with the kinds of local AI workloads he wanted to run and the performance he expected. He then set a budget of just over $2,500. He says he left the budget switched on deliberately. His point is that a VRAM-heavy request with no spending cap could drift back into RTX 5090 territory.

His planning assumption is that capable home models need at least a 16GB card, to get decent quantization and a workable context window. That is his workload-specific assumption, not a universal rule. If you plan to run bigger models, your threshold will differ.

The parts list XDB produced​

The planner landed on a $2,304 build, leaving $196 of the budget unspent. These are the prices XDA reported on publication day, not verified live quotes.

Two details stand out. The first is memory: the 32GB kit, at $480, costs more than the CPU. The author flags this himself. It is the obvious place to shop around or ask XDB for alternatives.

The second is the date. XDB's site says prices can change or be removed. Treat these numbers as a snapshot from November 5, 2026, not durable guidance.

Why the RX 9070 XT, and what the specs confirm​

The GPU is the biggest line item. The author argues the 16GB RX 9070 XT is one of the cheaper ways to get that much VRAM.

The headline figures check out against AMD's own listing. AMD lists 16 GB of GDDR6 on a 256-bit interface, with memory bandwidth of up to 640 GB/s. The same page gives a typical board power of 304W and a minimum PSU recommendation of 750W. It also lists two 8-pin power connectors and OS support for Windows 10 64-bit, Windows 11 64-bit and Linux x86 64-bit.

That power figure matters for the planner's choice of PSU. Check your own board's connector requirements and the PSU's rated capacity, particularly if you pick a different 9070 XT model. Partner designs vary. One retailer-style spec guide lists some ASUS 9070 XT models with three 8-pin connectors and 750W to 850W recommendations.

The article's comparison with the RX 9060 XT also holds. Wikipedia's RX 9000 series table lists 320 GB/s for the 9060 XT against 640 GB/s for the 9070 XT. The 448GB/s figure XDA quotes for the RTX 5060 Ti 16GB wasn't independently confirmed here, so treat it as XDA's claim.

Bandwidth and capacity are specifications, not benchmarks. XDA's test shows no tokens-per-second results, and neither does this article. Third-party estimators exist, but they are estimates. Before you buy, check what your target models need at your chosen quantization and context length.

The software caveat: ROCm, Windows and Linux are not the same​

XDB attached a note to the card saying the XFX Swift runs local AI tools through ROCm. It says Ollama, llama.cpp, LM Studio, vLLM and ComfyUI list it as supported, and that some workflows are tuned for Nvidia first. The warning about CUDA-first tooling is fair. The supported-everywhere impression needs qualifying, because support depends on the operating system and the release.

AMD's ROCm 7.1.1 Linux compatibility page does list the RX 9070 XT. It names Ubuntu 24.04.3, Ubuntu 22.04.5 and RHEL 10.1 as supported operating systems. It also pairs ROCm 7.1.1 with specific framework versions as official production support: PyTorch 2.9.1, ONNX Runtime 1.23.1, TensorFlow 2.20 and Triton 3.5.1.

Ollama's GPU documentation shows the split more sharply:

• Linux: the RX 9070 XT appears in the ROCm list, and Ollama requires the ROCm v7 driver stack.
• Windows: Ollama's Windows Radeon list shows RX 7000-series and PRO cards. It does not include the RX 9070 XT.
• Vulkan: Ollama separately documents Vulkan as an additional GPU route on Windows and Linux. It is enabled by default when the backend is installed.

The author notes that llama.cpp also supports Vulkan, and says Vulkan on RDNA 4 often generates tokens faster than ROCm. That is his observation, not something this article verified. Ollama's Vulkan note offers one practical pointer: without extra capabilities or root access, Ollama estimates model sizes when scheduling, rather than using reported VRAM data.

If you plan to run this build on Windows, check your chosen tool's current supported-GPU list, backend and driver requirements before buying. A card listed on one OS or release may not be listed on another.

The Nvidia switch and the used-card route​

XDB offers a one-click Nvidia swap. In the author's run it replaced the Radeon with an Asus TUF Gaming RTX 4080, which cost $794 more. That raised the total to $3,098, well over the budget.

The reported trade is narrow. Both cards have 16GB, so the 4080 won't hold larger models. It offers roughly 12% more bandwidth (716.8GB/s against 640GB/s) and CUDA compatibility. The author explains that the swap simply finds the cheapest Nvidia equivalent the planner can see, which was the 4080 at the time.

A used card is the other common change. In the author's test, putting an RTX 3090 into the planner showed no retailer price, so the price came up as unavailable. His workaround is to remove the card and handle the price math yourself, or leave the GPU slot empty if you already know what you're buying. He describes the interface steps as clicking "Manage" and then "-" to remove a card, or "Change" or "Manage" to swap parts and prompt for requirements such as a motherboard with Wi-Fi. Those steps come from his hands-on report. XDB's public page doesn't confirm them.

How to use a planner like this sensibly​

1. Set a budget deliberately. With it off, the planner follows your request alone, and a VRAM-heavy request can balloon.
2. State the workload. Say what models, quantization and tools you want to run, and on which OS.
3. Check the GPU for your software stack. Match the card, OS and backend against the tool's current documentation.
4. Review the cheap and the pricey outliers. The $480 memory kit is the obvious one in this build.
5. Verify the power budget. Confirm the PSU capacity and connectors against your specific GPU model.
6. Re-check prices before buying. Retail prices can move or vanish quickly.

What this does and doesn't show​

This is a planner demonstration. The author didn't assemble the machine or benchmark it, and neither this article nor its sources measured AI performance. The headline framing about RTX 5090 money is reported as a budget idea. XDA's claim that the build costs about a quarter of the 5090's current price rests on a price that wasn't independently sourced here, so treat that ratio with caution.

What the test does show is how quickly a tool can turn a rough idea into a fit-checked parts list. It also shows how easily a $200 GPU swap, a pricey memory kit, or an OS-specific driver gap can undo a budget. For a first local AI build, a planner is a decent first draft. The compatibility checking on your chosen software is still your job.