llm4funAI HARDWARE, EXPLAINED.
THE WORKBENCH
THE LOCAL AI SIMULATOR

What hardware can run your LLM?

Choose a model. Compare memory fit, estimated tokens per second and hardware prices for your budget.

104 hardware profilesOne workload. The whole picture.

The performance landscape

Find a setup by name, or select a point to highlight its price, speed and memory.

SIMULATED

Include results that use the CPU and system RAM.

5 010,02 USD ceiling42 setups above this range · search still covers all hardware

Generation speed · tok/s per userFull context · central estimates
NVIDIA consumerNVIDIA ProNVIDIA datacenterAppleAMD / IntelDGX Spark / ThorStrix HaloMulti-GPU / custom
Hollow markers = CPU offload required

↑ Faster · ← Lower cost

Select a point to highlight it, or search by name above.

● Single GPU■ Unified memory▲ Multi-GPULarger = more AI memoryFaded = over budget
Plot coverage · 61 shown / 58 omitted

16 systems have missing or unplottable values for these axes.

42 setups are above the 5 010,02 USD graph ceiling.

Search above or use Browse every setup below to explore the full hardware library.

Show
Speed and target matches are estimates; ranges show uncertainty, not a guarantee.

Purchase estimates use the selected market; missing local prices use labeled FX estimates from another market. The price graph defaults to a ceiling of 125% of your budget. Use the full price range or search to explore more expensive setups. GPU prices include the configured host PC. Used prices where available; otherwise new.

YOUR HARDWARE SHORTLIST

Different strengths. A good fit.

Every pick fits your model, meets 30 tokens/s per user, and stays within budget.

The budget pickFits your plan
UNIFIED SYSTEM

Complete computer · Unified memory

Mac mini M5 Pro (15-core CPU, 16-core GPU) 24GB

1 699USD

New purchase estimate

Estimated generation31–55tokens/s
Inference power100W
Memory fit6.7 GiB / 16 GiB usable

9.3 GiB of spare memory

The lowest priced match. Enough memory and estimated speed for this workload, with more of your budget left over.

The power-conscious pickFits your plan
UNIFIED SYSTEM

Complete computer · Unified memory

Mac Studio M4 Max (32-core GPU) 36GB

2 000USD

Used purchase estimate

Estimated generation34–61tokens/s
Inference power100W
Memory fit6.7 GiB / 24 GiB usable

17.3 GiB of spare memory

An estimated 100 W during inference. About 2 USD/month at 4 hours a day. A useful option when ongoing energy use matters.

The performance pickFits your plan
GPU SETUP

Single GPU · Host included

RTX 5080 16GB

3 000USD

Used purchase estimate

Estimated generation77–140tokens/s
Inference power340W
Memory fit7.0 GiB / 15 GiB usable

8.4 GiB of spare memory

Around 4.0× your speed target at the central estimate. A performance-focused option for more responsive generation.

A FEEL FOR THE SPEED

What does fast actually feel like?

Watch the same answer arrive at your target speed and on a system that fits your workload.

CURRENT WORKLOADQwen3 8B1 person · 8,192 context
YOUR PACE

Your target

30 tok/s

About 4.2 s for this ~125-token response.

Prompt

Write a short passage about a neighborhood garden.

Press play to see this response take shape.

Requested pace 30 tok/s

ELIGIBLE HARDWARE

RTX 5080 16GB

77–140 tok/s

About 1.1 s for this ~125-token response · 3 000 USD

Prompt

Write a short passage about a neighborhood garden.

Press play to see this response take shape.

Central estimate 118.8 tok/s

Illustrative preview at the central estimate. About 125 tokens; excludes prompt processing. Actual tokenization and speed vary.

KEEP EXPLORING

There’s more under the hood.

Already have a machine?Compare models on the hardware you own.Get comfortable with the numbers.Memory, quantization and speed, explained.
A clearer starting point

Plan a local AI setup around your workload

Compare local language-model hardware by memory fit, estimated generation speed and price. GPU VRAM, Apple-style unified memory and CPU offload behave differently, so the same model and settings can produce very different trade-offs. Adjust the workload and budget to explore a practical shortlist.

Questions people ask

What does this local AI hardware calculator estimate?

It estimates whether a selected model, quantization and context can run on each listed system, then compares a range of per-user generation speeds. It also shows hardware prices where available. These are planning estimates based on hardware and software assumptions, not measured benchmarks or a promise that a particular setup will perform the same in your environment.

Is GPU VRAM the same as unified memory or system RAM?

No. A discrete GPU usually has its own VRAM, while unified-memory systems let the processor and GPU share one memory pool. Ordinary system RAM can also hold model data, but moving it to a GPU over PCIe can limit speed. The simulator treats those memory paths differently; installed RAM plus VRAM should not be read as one equally fast pool.

Why compare speed ranges instead of one tokens-per-second number?

Generation speed varies with backend, drivers, batch size, model implementation and other workload details. A range communicates that uncertainty more honestly than a single precise-looking number. The estimate describes decoding tokens per second per person; prompt processing and time to first token are separate parts of an end-to-end response. Check how fit and speed are estimated and the reported inference measurements before buying hardware.

How much VRAM does an 8B LLM need?

It depends on the exact model, precision and context. For Qwen/Qwen3-8B converted to GGUF Q4_K_M, the simulator estimates 4.6 GiB for weights and 7.0 GiB in total with one user, an 8,192-token context, F16 KV cache, single-GPU llama.cpp CUDA buffers with flash attention and 5% headroom. Other 8B models can have different cache and runtime requirements. This is a planning estimate, not a measured allocation or a speed guarantee. Sources: Qwen3-8B official configuration.

Can one RTX 5090 run a 70B model?

An RTX 5090 has 32 GB of VRAM; llm4fun budgets 31.4 GiB as usable after its driver reservation. For meta-llama/Llama-3.3-70B-Instruct converted to GGUF Q4_K_M, the estimated weights alone take 39.7 GiB at 4.83 bits per weight. With one user, an 8,192-token context, F16 KV cache, llama.cpp CUDA runtime buffers with flash attention and 5% headroom, the total estimate is 45.3 GiB, so this configuration cannot fit fully on one card. CPU offload or a different quantization changes the memory and speed trade-offs; this is a planning calculation, not a measured run. Sources: NVIDIA RTX 5090 specifications; Meta Llama 3.3 model card.