Your next build.
Pick your parts. Keep the whole build in view. Check fit and retailer prices before you buy.
Loading build inventory…
Pick your parts. Keep the whole build in view. Check fit and retailer prices before you buy.
Loading build inventory…
Current starting build: Custom
Choosing another starting build loads its default parts.
Change a part in place. Check its price and buy from the same list.
Some links may earn us a commission.
Choose parts to see their compatibility checks.
System RAM: —
Estimated: CPU TDP + every GPU board power × purchase quantity + 75 W platform. Unknown powers are excluded from the partial subtotal. Not a benchmark.
The estimate needs known CPU and GPU power facts.
PLAN YOUR PLAY / LOCAL AI
Reviewed
4K text context · one session · 4 GiB planning reserve
Generation speed: Unmeasured
System RAM: —
Nominal GPU VRAM used for planning: —
≈ 3.3 GB download · ≈ 7.6 GB planning allocation
GPU memory unknown
Model source · Qwen3.5 4B≈ 6.6 GB download · ≈ 10.9 GB planning allocation
GPU memory unknown
Model source · Qwen3.5 9B≈ 8.0 GB download · ≈ 12.3 GB planning allocation
GPU memory unknown
Model source · Gemma 4 12B≈ 17 GB download · ≈ 21.3 GB planning allocation
GPU memory unknown
Model source · Qwen3.5 27BThis is a site planning heuristic. Actual memory depends on context, runtime, drivers and GPU support; a budget result does not confirm execution. System RAM is separate from VRAM.
Download sizes and planning allocations use decimal GB. The 4 GiB reserve is 4,294,967,296 bytes (≈ 4.3 GB) for runtime, context and desktop overhead, not an official memory requirement. Nominal GPU capacity is compared as decimal GB as a conservative convention; physical memory bytes are unmeasured.
Published sizes are rounded: ±0.05 GB for 3.3, 6.6 and 8.0; ±0.5 GB for 17. Within covers the entire allocation interval, Exceeds is below it, and Near overlaps it. Bundled model components may not all load for text inference; charging the full download does not establish actual VRAM use.
These exact Q4_K_M tags were checked October 11, 2026. No generation benchmark has been measured. Verify your backend, especially AMD support, and the actual processor split using ollama ps.
qwen3.5:4b-q4_K_M · digest prefix d8b0f5e9760cqwen3.5:9b-q4_K_M · digest prefix 56671c2ab938gemma4:12b-it-q4_K_M · digest prefix 6114515d63c1qwen3.5:27b-q4_K_M · digest prefix 64ae5b1d651fGPU references, not measurements of your complete build. CPU, RAM, scene, patches and drivers affect actual results.
Incomplete build context: choose a valid CPU and RAM before judging your complete setup.
Single-GPU published reference · primary GPU only. Additional GPUs do not multiply FPS.
Select a valid GPU to see published FPS references.
Publisher updated · checked
Ryzen 7 9800X3D at 5.4 GHz · ASUS X670E Hero · ReBAR enabled · Windows 11 · DDR5-6000 28-36-36-96 · Arctic Liquid Freezer III 420.
Raster tests use ray tracing off, with upscaling and dynamic resolution disabled unless noted. Wukong uses 100% scaling with FSR selected by the game, without upscaling. Chart headers take precedence over the methodology page, which notes a pending update.
No complete-build estimate, interpolation or CPU/RAM adjustment is applied. Unsupported or inconsistent GPU identities have no reference.
Original GPU charts · Publisher test methodology