GPUs and AI chips for local LLMs
Every card or chip with at least 2 builds or 4 measured results gets its own page with all its builds, results, memory, bandwidth and a dated EU price where we track one.
| GPU / chip | Memory | Bandwidth | Typical EU price | Builds | Results | Fastest plain decode |
|---|---|---|---|---|---|---|
| DGX Spark / GB10 | 128 GB unified memory per GB10 system | 9 | 33 | GLM-5.3 Flash: 105.3 | ||
| Strix Halo (Ryzen AI Max+ 395, Radeon 8060S) | unified memory: 96–128 GB in our builds | 256 GB/s | 9 | 32 | Llama 3.2 3B: 88.14 | |
| RTX 3090 | 24 GB per card | 936 GB/s | ≈€1,248 | 20 | 30 | gpt-oss 20B: 207.2 |
| RTX 4090 | 24 GB per card | 1008 GB/s | ≈€3,955 | 9 | 24 | Llama 3.2 3B: 334.67 |
| RTX 5090 | 32 GB per card | 1792 GB/s | ≈€7,140 | 8 | 20 | gpt-oss 20B: 324.2 |
| Radeon AI Pro R9700 | 32 GB per card | ≈€1,848 | 5 | 20 | Llama 3.2 3B: 195.91 | |
| Apple M3 Ultra | unified memory: 512 GB in our builds | 7 | 15 | Llama 3.2 3B: 154.6 | ||
| Apple M5 Max | unified memory: 36–128 GB in our builds | 3 | 14 | Qwen3.6 35B-A3B: 145.8 | ||
| RTX 3060 12GB | 12 GB per card | 360 GB/s | ≈€429 | 3 | 10 | Llama 3.2 3B: 122.85 |
| RTX Pro 6000 Blackwell | 96 GB per card | ≈€16,885 | 3 | 9 | gpt-oss 20B: 286.91 | |
| Apple M5 Ultra | unified memory | 3 | 9 | Qwen3.8 27B: 50.89 | ||
| RTX Spark N1X | unified memory | 2 | 8 | Qwen3.6 35B-A3B: 89.5 | ||
| Intel Arc Pro B70 | 32 GB per card | 3 | 8 | Gemma 4 26B-A4B: 122.16 | ||
| RTX 5070 Ti | 16 GB per card | 896 GB/s | ≈€1,245 | 3 | 6 | gpt-oss 20B: 189.45 |
| Radeon RX 7900 XTX | 24 GB per card | 1 | 6 | Qwen3.6 35B-A3B: 87.2 | ||
| RTX 5060 Ti 16GB | 16 GB per card | ≈€739 | 2 | 5 | gpt-oss 20B: 111.51 | |
| Tesla V100 32GB | 32 GB per card | 889 GB/s | 1 | 5 | Qwen3.8 27B: 42.4 | |
| Apple M1 Ultra | unified memory | 2 | 5 | Llama 3.2 3B: 127.6 | ||
| RTX 4060 Ti 16GB | 16 GB per card | 288 GB/s | 3 | 4 | LFM2.5 24B-A2B: 131-150 | |
| RTX 5080 | 16 GB per card | 960 GB/s | ≈€1,770 | 2 | 3 | gpt-oss 20B: 204.85 |
| RTX A4000 | 16 GB per card | ≈€1,050 | 2 | 3 | Llama 3.2 3B: 149.44 | |
| Apple M1 Max | unified memory | 2 | 2 | gpt-oss 20B: 75.15 |
Prices and bandwidth are shown only where sourced (see each page). “Fastest plain decode” excludes speculative decoding, pending and partly-offloaded results, and mixes models, so compare on the GPU pages.
Other hardware with fewer results
RTX 4070 Ti: Intel Core Ultra 265K + RTX 4070 Ti · Apple M2 Ultra: Mac Studio M2 Ultra 76-core GPU, 192 GB · Apple M4 Pro: MacBook Pro M4 Pro 48 GB: Flint 40B-A2.5B 4-bit at 127.9 tok/s (Kapitsa), 126.6 (rec), 79.6 (llama.cpp)
Also: best local LLMs by VRAM size · Can X run on Y? · free dataset