Skip to main content

Hardware Tiers

Device classification and chip family reference.

Device Tiers

TierNameExamplesRole
1BackboneMac Studio, Mac Pro, Linux servers with NVIDIA/AMD GPUAlways-on inference nodes. Serve large models, high throughput.
2DesktopMacBook Pro, Mac Mini, Linux/Windows desktops with GPUPrimary compute. Available when in use, may sleep.
3TabletiPad Pro (M-series), high-end Android tabletsLight inference. Small models only, intermittent availability.
4Phone/LeafiPhone, Android phones, single-board computersConsumer only. Requests inference but does not serve it.

Tier numbering is inverted: tier 1 is the highest capability. When filtering peers with minTier, a value of 2 means "tier 2 or better" (tiers 1 and 2).

Chip Family Reference

Apple Silicon -- Mac

ChipFamilyGenerationGPU CoresMax RAMBandwidthWattsTier
Apple M1m117-816 GB68 GB/s20W2
Apple M1 Prom1Pro114-1632 GB200 GB/s30W2
Apple M1 Maxm1Max124-3264 GB400 GB/s40W1-2
Apple M1 Ultram1Ultra148-64128 GB800 GB/s60W1
Apple M2m228-1024 GB100 GB/s22W2
Apple M2 Prom2Pro216-1932 GB200 GB/s35W2
Apple M2 Maxm2Max230-3896 GB400 GB/s45W1-2
Apple M2 Ultram2Ultra260-76192 GB800 GB/s65W1
Apple M3m338-1024 GB100 GB/s22W2
Apple M3 Prom3Pro314-1836 GB150 GB/s36W2
Apple M3 Maxm3Max330-40128 GB400 GB/s48W1-2
Apple M3 Ultram3Ultra360-80192 GB800 GB/s70W1
Apple M4m441032 GB120 GB/s22W2
Apple M4 Prom4Pro416-2048 GB273 GB/s38W2
Apple M4 Maxm4Max432-40128 GB546 GB/s50W1-2
Apple M4 Ultram4Ultra464-80256 GB819 GB/s75W1

Apple Silicon -- iPhone/iPad

ChipFamilyGenerationWattsTier
Apple A14 Bionica14145W4
Apple A15 Bionica15155W4
Apple A16 Bionica16166W4
Apple A17 Proa17Pro176W4
Apple A18a18187W4
Apple A18 Proa18Pro187W4
Apple A19 Proa19Pro198W4

iPads with M-series chips (M1, M2, M4) use the Mac chip families and are classified as tier 3.

Non-Apple

Chip TypeFamilyWattsTierGPU Backend
NVIDIA GPU (RTX 3090/4090 class)nvidiaGPU300W1cuda
AMD GPU (RX 7900 class)amdGPU250W1rocm
Intel CPU (desktop)intelCPU65W2cpu or sycl
AMD CPU (desktop)amdCPU65W2cpu
ARM Generic (Snapdragon, RPi)armGeneric10W3-4cpu or vulkan
Unknownunknown30W2cpu

RAM and Model Capacity

The available RAM for model loading is:

availableRAMForModelsGB = max(totalRAMGB - 4.0, 1.0)
Total RAMAvailable for ModelsLargest Model (q4)
8 GB4 GBLlama 3.2 3B (1.8 GB)
16 GB12 GBLlama 3.1 8B (4.5 GB) or Phi 4 14B (8 GB)
32 GB28 GBQwen 3 32B (18 GB)
48 GB44 GBQwen 3 32B (18 GB) + spare
64 GB60 GBLlama 4 Scout 109B MoE (56 GB)
128 GB124 GBMultiple large models simultaneously
192 GB188 GBAny model or combination

For discrete GPUs (gpuVRAMGB field), the VRAM limit determines maximum model size, independent of system RAM.