Local LLM Inference
Choose GPU memory and supporting parts for running language models locally.
Use Builder to check power, PCIe lanes, and parts compatibility.
Open the builder baseline →AI workstation planning engine
How much GPU memory do you need?
Choose a model and precision to estimate memory. Then check context length or plan the complete workstation.
Choose your parts
Builder & Calculator
Choose parts, check compatibility and power requirements, then review or share the build.
Build Configuration
Select parts, then check power and compatibility.
Planning direction only: confirm slot spacing, chassis clearance, cabling, and deployment constraints before procurement approval. Validation methodology
Memory evidence: No GPU-specific memory measurements published here.
Copy the current selections for review.
Component specifications
System Breakdown
Total GPU Watts: 450W
Total CPU Watts: 170W
Total VRAM: 24GB
RAM Watts: 32W
NVMe Watts: 7W
System Overhead: 75W
GPU
RTX 4090
CPU
Ryzen 9 9950X
RAM
Corsair Vengeance DDR5 128GB (4x32GB) 5600MHz
NVMe
Samsung 990 PRO
Motherboard
ASUS ROG Crosshair X670E Hero
Socket: AM5
Max GPUs: 2
PCIe Slots: 2
Slot Spacing: standard
Recommended PSU Models
Corsair HX1200 — 1200W
Corsair AX1600i — 1600W
Seasonic PRIME TX-1600 — 1600W
EVGA SuperNOVA 1600 P+ — 1600W
Corsair AX2000 — 2000W
Build Summary
Recommended PSU Target: 1100W
Estimated Wall Power: 899W
Estimated System Load: 809W
Compatibility status: required checks are passing, but one or more items need review.
Compatibility Status
- UNVERIFIED: ASUS ROG Crosshair X670E Hero does not yet encode a board-specific CPU power/core ceiling; verify the manufacturer CPU support list before procurement.
All 12 compatibility checks
- PASS: Ryzen 9 9950X and ASUS ROG Crosshair X670E Hero share the AM5 platform.
- PASS: Ryzen 9 9950X's CPU family is represented in ASUS ROG Crosshair X670E Hero's supported platform set.
- UNVERIFIED: ASUS ROG Crosshair X670E Hero does not yet encode a board-specific CPU power/core ceiling; verify the manufacturer CPU support list before procurement.
- PASS: ASUS ROG Crosshair X670E Hero supports the requested 1 CPU(s) at the socket-count level.
- PASS: RTX 4090 is a PCIe add-in accelerator.
- PASS: ASUS ROG Crosshair X670E Hero has cataloged expansion capacity for 1 PCIe GPU(s) with the current CPU count.
- PASS: RTX 4090's 3.5-slot catalog width does not trigger a blocking fitment rule for this planning configuration.
- PASS: 128 GB is within ASUS ROG Crosshair X670E Hero's cataloged 192 GB memory limit.
- PASS: Selected DDR5 memory matches ASUS ROG Crosshair X670E Hero.
- PASS: Selected UDIMM memory matches ASUS ROG Crosshair X670E Hero's supported memory class.
- PASS: Ryzen 9 9950X's 24-lane catalog budget meets the 16-lane x16-per-GPU planning target.
- PASS: The 1100W PSU target is covered by the current catalog (up to 2000W).
Builder Validation Notes
- Platform Expansion Headroom Caution: This configuration is closer to the edge of platform expansion comfort. A build can be planning-level supported while still leaving less room for future growth than expected.
Manual validation still required: physical fit, cable/connector readiness, airflow/thermal density, BIOS or vendor specifics, and deployment environment constraints.
PCIe Lane Status
CPU lane budget: 24 • Minimum target: 8 • Preferred target: 16
- PCIe lane allocation looks acceptable for the selected GPU count.
CPU lane reference: AMD Ryzen 7000/9000 desktop specs (24 total PCIe lanes)
GPU Fitment Advisory
- Selected GPU width: 3.5 slots.
- Requested GPUs: 1.
- Estimated slot usage: 3.5 slots (typical baseline: ~7 slots).
- Selected motherboard uses standard slot spacing.
- Single-GPU fitment is typically straightforward; clearances still depend on exact cooler dimensions and chassis.
Why this board for this build?
- Matches the selected CPU platform.
- Supports the requested PCIe GPU count at the catalog level.
- A practical fit for lower-GPU-count workstation configurations.
- Provides PCIe slot capacity aligned with the requested build size.
Recommended next step
- Confirm fitment and lane notes above.
- Share this build URL for internal review.
- Validate current market pricing and availability before purchase.
Need deeper context? Cross-check compare tools, recommended baselines, and methodology notes.
Start here
Choose hardware for your workload.
Use the guides below to compare GPUs, choose a platform, or review a starting build.
Start with the task you want to run.
Image + Video Generation
Compare GPUs for image and video generation, including memory, power, and cooling needs.
Check GPU specifications before choosing parts.
Compare GPU options →GPU Memory Capacity
Compare 24GB, 48GB, and 96GB GPU memory tiers for local models.
Check model size and context needs before choosing a memory tier.
Review VRAM tier guide →Multiple GPUs
Check when a second GPU is useful and what additional GPUs require from the system.
Review software support, PCIe lanes, power, and cooling before adding GPUs.
Read GPU-count planning guide →Desktop, Workstation, or Server
Compare desktop, workstation, and server platforms for your workload.
Check expansion, memory support, and deployment requirements.
Review platform choice guide →Starting Builds
Review starting configurations and their compatibility limits.
Open a supported preset in Builder to check or change its parts.
Browse recommended builds →Recommended baselines
Recommended AI Workstation Builds
Choose a starting configuration. Presets with passing required compatibility checks can open in Builder; other examples remain reference-only. Check physical fit, cooling and cabling before buying.
Builder-validated
Creator AI Rig
Balanced single-GPU workstation for content generation, local assistants, and accelerated creative workflows.
Planning details
Why this build: Designed for high-VRAM creator workflows where local iteration is the priority rather than rack-scale deployment.
Best for:
- Stable Diffusion users and AI artists
- Solo creators building local copilots
- Developers prototyping 7B–13B local LLM apps
Performance:
- Single-GPU planning baseline for image generation, video enhancement, and local-assistant workflows
- Planning fit for local 7B–13B-class inference and RAG development; actual responsiveness depends on model, quantization, and software stack
- Combines creator workloads, upscaling, and local AI tooling on one workstation-class path
Upgrade path: Move to a dual-GPU motherboard platform or increase NVMe capacity for larger datasets and checkpoint libraries.
GPU Configuration: 1 × RTX 4090
CPU: 1 × Ryzen 9 9950X
Use Case: Image/video generation, RAG apps, and daily local inference development.
Open exact build in Builder →Builder-validated
Local Inference Workstation
Single-GPU workstation plan for local model serving, retrieval and development. Case, cooling and acoustic performance still require validation.
Planning details
Why this build: Pairs ample VRAM with a high-core desktop CPU so teams can run always-on local model services without datacenter complexity.
Best for:
- Product teams shipping internal copilots
- Developers testing RAG chains against private data
- Practitioners running 13B-class models continuously
Performance:
- Single-GPU planning baseline for local assistant and API-like inference workloads
- CPU and storage capacity are allocated for embedding, retrieval, and indexing alongside inference
- Workstation path for teams evaluating whether multi-GPU serving is actually required
Upgrade path: Add a second workstation GPU on a workstation platform when concurrent users or multi-model serving increases.
GPU Configuration: 1 × RTX 5000 Ada
CPU: 1 × Ryzen 9 9950X
Use Case: Local model serving, RAG development, and internal assistant prototyping.
Open exact build in Builder →Builder-validated
LoRA Fine-Tuning Workstation
High-VRAM dual-GPU planning baseline for parameter-efficient fine-tuning and medium-scale training workflows.
Planning details
Why this build: Designed to balance workstation ergonomics, GPU memory, and CPU resources for repeated LoRA and QLoRA fine-tuning cycles.
Best for:
- ML engineers running LoRA and QLoRA experiments
- Teams validating model adaptation before cloud scale-out
- Practitioners processing medium-sized private datasets locally
Performance:
- Dual-GPU layout provides two accelerators for parallel experiment scheduling
- Workstation CPU resources support preprocessing and tokenization alongside GPU jobs
- Planning baseline for LoRA/QLoRA tuning and evaluation; actual throughput depends on model, precision, batch size, and software stack
Upgrade path: Scale to four GPUs on the same platform and expand system RAM for larger batch sizes and concurrent jobs.
GPU Configuration: 2 × RTX 6000 Ada
CPU: 1 × Threadripper PRO 7975WX
Use Case: LoRA/QLoRA fine-tuning, quantization experiments, and heavier data preprocessing.
Open exact build in Builder →Reference architecture
Dual-GPU Research Workstation
Balanced research node for teams comparing model behavior, evaluation pipelines, and multi-run experiments.
Planning details
Why this build: Delivers the memory and PCIe capacity needed for serious experimentation while staying easier to deploy than 4-GPU platforms.
Best for:
- ML researchers running side-by-side model tests
- Teams validating prompt, adapter, and retrieval strategies
- Applied AI groups requiring reproducible local benchmarks
Performance:
- Dual 96GB GPU configuration provides 192GB aggregate VRAM for workloads that can distribute across devices
- Workstation CPU resources support preprocessing, orchestration, and metrics pipelines
- Planning baseline for recurring local evaluation and model-comparison workflows
Upgrade path: Expand to four GPUs on the same WRX90 platform when experiment concurrency requirements rise.
GPU Configuration: 2 × RTX PRO 6000 Blackwell Workstation Edition
CPU: 1 × Threadripper PRO 7975WX
Use Case: Model evaluation, dual-stream inference, and long-context analysis.
Builder handoff intentionally unavailable.
2450W planning target exceeds the current 2000W single-PSU Builder model.
Review reference architecture →