Published · Phase 3
Choose Hardware for a Local LLM
A useful LLM hardware calculator should recommend a tier from model size, precision, context, concurrency, latency, and budget—not merely say whether weights fit. Use the result as a shortlist, then validate the top configuration with a representative run.
Editorial review: complete · Updated 2026-08-30
Direct answer
Direct answer guidance: Recommend RTX 5090 below ~28 GB total, 48 GB below ~42 GB, RTX PRO/80–96 GB below ~85 GB, DGX Spark below ~105 GB, and multi-GPU/cloud above that—subject to latency and support needs.
Model / quant variants
Model / quant variants guidance: Select llama.cpp for GGUF/local flexibility and vLLM for concurrent serving. Runtime choice can change memory and throughput enough to alter the hardware recommendation.
Approximate weight footprint
Approximate weight footprint guidance: Calculate weight bytes from parameters and precision, add 10–20% runtime overhead, estimate KV cache for context × sequences, add multimodal/adapters/draft models, then require 10–20% spare capacity.
Hardware tiers
Hardware tiers guidance: Calculate weight bytes from parameters and precision, add 10–20% runtime overhead, estimate KV cache for context × sequences, add multimodal/adapters/draft models, then require 10–20% spare capacity. If the recommendation is too expensive, reduce precision/context, choose a smaller model, use retrieval, or rent cloud hardware for occasional jobs.
Observed configurations
Observed configurations guidance: Estimate speed from the closest public model/hardware/runtime results, then publish a range widened for differences in precision, prompt, and batching. Approximation is useful when the uncertainty is explicit.
Context / concurrency effects
Context / concurrency effects guidance: Make context and concurrent sequences explicit inputs. Using a model's maximum context by default will systematically overstate hardware needs and understate latency.
Performance expectations
Performance expectations guidance: Estimate speed from the closest public model/hardware/runtime results, then publish a range widened for differences in precision, prompt, and batching. Approximation is useful when the uncertainty is explicit.
Best-value configurations
Best-value configurations guidance: Compare capital cost, energy, support, utilization, cloud/API alternative, and operator time. A cheaper system that misses latency or fit requirements is not lower cost. Recommend RTX 5090 below ~28 GB total, 48 GB below ~42 GB, RTX PRO/80–96 GB below ~85 GB, DGX Spark below ~105 GB, and multi-GPU/cloud above that—subject to latency and support needs.
What will not fit / weak evidence
- Display estimates as ranges and list the assumptions that would move the recommendation up or down a tier.
- Recompute when the model revision, quantization, runtime, context, or concurrency changes.
Planner CTA
Compare this recommendation against your model, context, concurrency, latency, and budget in the ComputeSage Planner.
Sources / methodology
Recommendations combine official specifications, public model/runtime documentation, adjacent benchmark observations, and clearly labeled engineering estimates. Review the ComputeSage evidence methodology.
- NVIDIA GeForce RTX 5090 specifications
- NVIDIA RTX PRO 6000 Blackwell specifications
- NVIDIA DGX Spark specifications
- Qwen3.8-27B model repository
- Qwen3.6-35B-A3B model card
- OpenAI gpt-oss introduction and deployment guidance
- Mistral Small 4 model card
- DeepSeek-V4-Flash model card
- vLLM installation and hardware documentation
- llama.cpp project and backend documentation
- StackBench controlled lab approved source qwen38-rtx5090-owner-matrix-20260816