GPU for AI Buying Advisor

Compare memory fit and your own GPU quotes. A shortlist, not a compatibility promise.

Saved only in this browser

Your model and budget

Start with the model, context and budget. Replace sample quotes before buying.

GPU budget ($)

Reference model / cache geometryNominal parameters; full-attention geometry from the linked model config.

Model parameters (billions)

Weight precision

Context tokens per session

Qwen2.5 32B nominal weight scenario; 64 layers, 40 attention heads, 8 KV heads, head dimension 128.

Quotes and usable VRAMReplace sample prices and usable memory budgets. Disabled quotes are excluded.

Labels retain vendor capacity names in GB. The numeric budgets below are editable GiB scenarios, not measured free VRAM or GB-to-GiB conversions. Enter the usable capacity reported by your runtime; reserve and safety are additional allowances. No multi-GPU pooling.

RTX 3060 12GBInclude RTX 3060 12GB

RTX 3060 12GB quote ($)

RTX 3060 12GB usable VRAM (GiB)

RTX 5080Include RTX 5080

RTX 5080 quote ($)

RTX 5080 usable VRAM (GiB)

RTX 4090Include RTX 4090

RTX 4090 quote ($)

RTX 4090 usable VRAM (GiB)

RTX 5090Include RTX 5090

RTX 5090 quote ($)

RTX 5090 usable VRAM (GiB)

Custom single GPUInclude Custom single GPU

Custom single GPU quote ($)

Custom single GPU usable VRAM (GiB)

Memory allowancesThe same byte-based model as the Local AI Hardware Checker.

Concurrent sessions

Runtime reserve (GiB)

Safety allowance (%)

Weight format allowance (%)

KV cache precision

Try an exampleIllustrative inputs, not a measured machine or a current quote.

Estimated full-GPU memory

22.84 GiB

RTX 4090 — estimated fit, tight

Lowest entered price among enabled in-budget memory fits: $1,600.00. 1.16 GiB headroom. Memory fit alone does not establish speed, compatibility or the best purchase.

Weight payload: 14.90 GiB
Weight format allowance: 1.86 GiB
Context / KV cache: 2.00 GiB
Runtime reserve: 2.00 GiB
Safety allowance: 2.08 GiB

Your enabled shortlist

RTX 3060 12GB$250.0010.84 GiB over budget · over memory budget · within purchase budget
RTX 5080$999.006.84 GiB over budget · over memory budget · within purchase budget
RTX 4090$1,600.001.16 GiB headroom · estimated fit, tight · within purchase budget
RTX 5090$1,999.009.16 GiB headroom · estimated fit · within purchase budget
Custom single GPU—Excluded — enable only for a real single-GPU quote.

Tight means less than max(1 GiB, 10% of the estimate) remains, assessed before rounding. This screening threshold is not a reliability guarantee.

Only this deliberate action puts these memory assumptions, prices and quote choices into a share link. Legacy links without a geometry use the default Qwen2.5 32B geometry; check the profile before relying on the result.

Calculated in this browser. Inputs are not saved automatically or sent to analytics.

Method, assumptions and limits

Required bytes = weight payload + format allowance + dense KV cache + runtime reserve, then the entered safety percentage on that subtotal. All displayed memory estimates use GiB (2³⁰ bytes). One billion parameters means 10⁹ parameters. System RAM is not added to VRAM.

Q4 starts at 4 bits plus a 12.5% format allowance; Q8 starts at 8 bits plus 6.25%. These are approximate weight scenarios, not exact Q4_K_M tensor residency. Quantized cache scenarios include scales and complete-row block padding. They do not establish backend support. Sessions share one weight instance, not separate model processes.

The default Qwen2.5 32B example uses 64 layers, 8 KV heads and a 128-value head dimension. At 8,192 tokens its FP16 cache is 2 GiB. With 2 GiB runtime reserve and 10% safety, the estimate is 22.84 GiB. A 24 GiB usable budget is an estimated tight fit, rather than the former mixed-unit 24.2 GB rejection.

Sources and purchase boundary

Geometry and quantization sources reviewed 2026-09-11: Qwen2.5 32B official config, Meta Llama 3 architecture, ggml block format definitions. The Hardware Checker exposes the other cited profiles and custom geometry.

Vendor capacity labels: RTX 3060, RTX 5080, RTX 4090 and RTX 5090. Prices are illustrative user-editable numbers, not vendor quotes, deals or availability claims.

Compare returns, warranty, cooling, power supply, software support and actual benchmarks before buying. This model excludes loading peaks, runtime-specific offload, multi-GPU sharding, training, multimodal, MoE, hybrid and sliding-window caches. It does not predict tokens per second. For full-system power, use the electricity calculator rather than interpreting GPU rated power as a measured bill.

FAQ

Does this guarantee that a model will run?

No. It screens estimated memory against a user-entered usable budget. Runtime, model-file, loading and performance checks remain separate.

Does more system RAM or another GPU automatically count?

No. The shortlist compares a full-model estimate against one GPU budget at a time. It does not verify sharding or offload allocation.

Why can a 24 GiB budget now show a tight fit?

The advisor uses consistent bytes and the cited Qwen geometry. Its default estimate is about 22.84 GiB, not the previous mixed-unit 24.2 GB. Tight is a screening result, not a certified runtime fit.