How much VRAM?
Separates the model-weight floor from runtime planning headroom so minimum fit is not confused with a safer deployment target.
AI Hardware Calculator: How Much GPU Do I Need?
Choose a model, precision or quantization level, workload size, and goal. ComputeAtlas estimates the VRAM, GPU count, system RAM, storage, and system class you should plan around before buying a local AI workstation or infrastructure node.
Estimates use the shared ComputeAtlas workload reference and distinguish model-weight floors from runtime planning headroom. Fine-tuning estimates represent parameter-efficient LoRA/QLoRA-style adaptation, not full-parameter training. See the VRAM methodology and source reference.
What this calculator answers
Turn a model choice into a hardware planning envelope
How many GPUs?
Converts the selected workload into a GPU-count direction while keeping unsupported or catalog-limited outcomes explicit.
How much RAM and storage?
Sizes the supporting system around the workload instead of treating the GPU as the only capacity decision.
Can I open a compatible build?
A Builder preset is exposed only when every required compatibility check passes; otherwise the result stays reference-only.
Current model references: Llama 3 8B, Llama 3 70B, Mixtral 8x7B, DeepSeek LLM 67B, Qwen3 8B, Qwen3 32B, Mistral Small 3.1 24B, Gemma 3 27B, DeepSeek-R1-Distill-Qwen-32B, DeepSeek-R1 671B, gpt-oss-20b, gpt-oss-120b, SDXL 1.0, FLUX.1 dev. The supported matrix uses one shared workload methodology across inference and parameter-efficient fine-tuning.
Decision visualThe estimator converts workload assumptions into a system envelope
VRAM target18 GB
GPU count1
System RAM64 GB
Trust & provenance — how to read these recommendations
See which parts of this estimate come from published model facts, deterministic derivation, and ComputeAtlas planning methodology.
Hardware evidence. The registry currently contains 21 manufacturer-source evidence records, with 8 of 8 current motherboard identities represented. Evidence records verify specific fields only; unresolved planning fields are not promoted to manufacturer claims.
Compatibility contract. Builder-validated status requires every required compatibility check to pass. Advisory warnings and unverified checks may remain explicit and non-blocking, and physical/OEM validation is still required before procurement.
Recommendation claims. The current recommendation set contains 42 planning-guidance claims and 0 benchmark-evidenced claims. Quantitative throughput or latency claims require a reproducible benchmark protocol before they can be classified as benchmark-evidenced. No benchmark evidence records are currently published.
Estimator methodology. workload-methodology-v1 (v1, effective 2026-09-18) distinguishes source-backed model facts, deterministic model-weight derivation, and ComputeAtlas planning assumptions for runtime reserve, GPU count, system RAM, storage, and build class.
Still requires final validation. Chassis fit, slot geometry, power connectors, thermals, BIOS/firmware, OEM topology, and deployment-specific constraints remain outside a generic planning guarantee.
Read the full ComputeAtlas methodology or review the model-memory source reference.
Estimator Inputs
Set your assumptions, then generate conservative sizing guidance for early planning and procurement discussions.
Estimated Requirements
Directional System Class
Single-Node Inference Build
Prioritizes reliable single-node inference capacity for initial deployment.
- Estimated VRAM required
- 18 GB
- Recommended GPU tier
- Prosumer (24 GB class)
- Recommended GPU count
- 1 GPU
- Recommended system RAM
- 64 GB
- Recommended storage tier
- 1 TB NVMe
- Suggested build class
- Single-Node Inference Build
Inference planning includes a runtime reserve above the raw model-weight floor; actual KV cache, batch, framework, and context costs vary by runtime.
Why this result
- Llama 3 8B at FP16 has an approximate 16 GB model-weight floor and a 18 GB standard inference planning target before workload scaling.
- Standard workload and inference assumptions produce a 18 GB planning target with runtime headroom rather than a minimum-fit claim.
- 1 GPU and 1 TB NVMe indicate a single-node inference build profile.
What moves you up or down a tier
- Move up a tier if you expect larger models, higher concurrent request load, longer context, or want additional VRAM headroom for reliability.
- Move down a tier only when workload shape is stable (smaller models or lower precision) and utilization targets are intentionally constrained.
Open Compatible Preset in Builder
This preset passes every required compatibility check currently encoded by ComputeAtlas, meets or exceeds the displayed RAM target, and stays within the current PSU catalog ceiling. Final physical, OEM, thermal, and deployment-specific validation is still required.