Hardware guide
Compare machines before choosing a local AI path
This table compares common PCs, GPUs, unified-memory Macs and datacenter systems for local AI work. Use it to estimate which machines are realistic for learning, single-user experiments, multi-session serving or large-scale inference.
Memory capacity decides what can fit. Memory bandwidth is a rough signal for generation speed, but real LLM performance also depends on the runtime, quantization, context length and concurrency.
Before buying hardware: verify the selected hardware against the models, runtime, quantization, context length and concurrency you plan to use. Treat these estimates as a starting point, then check recent hands-on tests, benchmarks and videos for that setup.
| Local AI hardware reference list | ||||
|---|---|---|---|---|
| Hardware | Memory | Bandwidth | Computing power | vLLM |
| RTX 3060 12GBNVIDIA GeForce RTX 3060 12 GB desktop GPU | 12 GBGPU VRAM | 360 GB/s | ~101 TOPS | yes |
| DDR4 16GB ComputerComputer with 16 GB DDR4-3200 dual-channel RAM | 16 GBRAM | 51.2 GB/s | ~2 TOPS | no |
| DDR5 16GB ComputerComputer with 16 GB DDR5-5600 dual-channel RAM | 16 GBRAM | 89.6 GB/s | ~3 TOPS | no |
| RTX 2000 AdaNVIDIA RTX 2000 Ada Generation 16 GB | 16 GBGPU VRAM | 224 GB/s | ~191.9 TOPS | yes |
| RTX 4060 Ti 16GBLow-cost NVIDIA 16 GB VRAM desktop GPU baseline | 16 GBGPU VRAM | 288 GB/s | ~353 TOPS | yes |
| RTX 5060 Ti 16GBCurrent low-cost NVIDIA 16 GB VRAM desktop GPU class | 16 GBGPU VRAM | 448 GB/s | ~759 TOPS | yes |
| RTX 5080NVIDIA GeForce RTX 5080 16 GB | 16 GBGPU VRAM | 960 GB/s | ~1801 TOPS | yes |
| Mac mini M4Apple Mac mini M4 system with 24 GB unified memory | 24 GBUnified RAM | 120 GB/s | ~40 TOPS | no |
| M4 ProApple M4 Pro with 24 GB unified memory | 24 GBUnified RAM | 273 GB/s | not available | no |
| RTX 3090 TiNVIDIA GeForce RTX 3090 Ti 24 GB | 24 GBGPU VRAM | 1008 GB/s | ~320 TOPS | yes |
| DDR5 32GB ComputerComputer with 32 GB DDR5-5600 dual-channel RAM | 32 GBRAM | 89.6 GB/s | ~3 TOPS | no |
| RTX 5090NVIDIA GeForce RTX 5090 32 GB desktop GPU class | 32 GBGPU VRAM | 1792 GB/s | ~3352 TOPS | yes |
| M5 ProApple MacBook Pro M5 Pro 20-core GPU configuration with large unified memory | 64 GBUnified RAM | 307 GB/s | ~180 TOPS | no |
| Radeon 8060S 96GBAMD Radeon 8060S Graphics with 96 GB allocated unified memory | 96 GBUnified RAM | 256 GB/s | not available | yes |
| M3 UltraApple Mac Studio M3 Ultra 80-core GPU configuration with 96 GB unified memory | 96 GBUnified RAM | 819 GB/s | ~120 TOPS | no |
| RTX PRO 6000NVIDIA RTX PRO 6000 Blackwell series | 96 GBGPU VRAM | 1792 GB/s | ~3511 TOPS | yes |
| DGX SparkNVIDIA DGX Spark desktop AI system with unified memory | 128 GBUnified RAM | 273 GB/s | ~1000 TOPS | yes |
| M5 MaxApple MacBook Pro M5 Max 40-core GPU configuration | 128 GBUnified RAM | 614 GB/s | ~320 TOPS | no |
| M5 UltraApple Mac Studio M5 Ultra 80-core GPU configuration with 256 GB unified | 256 GBUnified RAM | 1200 GB/s | ~540 TOPS | no |
| DGX StationNVIDIA DGX Station with GB300 Grace Blackwell Ultra | 748 GBCoherent Memory | 7100 GB/s | ~20000 TOPS | yes |
| ET900N G3ASUS ExpertCenter Pro ET900N G3 with NVIDIA GB300 Grace Blackwell Ultra | 748 GBCoherent Memory | 7100 GB/s | ~20000 TOPS | yes |
| DGX H200NVIDIA 8-GPU Hopper rackmount AI system with 1,128 GB total HBM3e, 141 GB per GPU and 4.8 TB/s HBM bandwidth per GPU | 1128 GBGPU HBM3e | 4800 GB/s | ~3958 TOPS | yes |
| DGX B200NVIDIA 8-GPU Blackwell rackmount AI system with 1,440 GB total HBM3e, 180 GB per GPU and 64 TB/s aggregate HBM bandwidth | 1440 GBGPU HBM3e | 8000 GB/s | ~18000 TOPS | yes |
| GB200 NVL72NVIDIA rack-scale liquid-cooled system with 72 Blackwell GPUs, 36 Grace CPUs, 13.4 TB total HBM3e, 576 TB/s aggregate HBM bandwidth and 130 TB/s aggregate NVLink | 13400 GBGPU HBM3e | 8000 GB/s | ~20000 TOPS | yes |
| GB300 NVL72NVIDIA rack-scale liquid-cooled system with 72 Blackwell Ultra GPUs, 36 Grace CPUs, about 20 TB total HBM3e, up to 576 TB/s aggregate HBM bandwidth and 130 TB/s aggregate NVLink | 20000 GBGPU HBM3e | 8000 GB/s | ~20000 TOPS | yes |