Apple M2 Ultra
apple Apple Silicon (M2 Ultra, unified memory) 2023
Fields
id
accelerators/apple-m2-ultraVendor
apple
Architecture
Apple Silicon (M2 Ultra, unified memory)
Release year
2023
Process (nm)
5
Form factors
integrateddesktop
VRAM (GB)
192
Memory type
LPDDR5 (unified)
Memory bus (bit)
1,024
Memory bandwidth (GB/s)
800
Bandwidth basis
M2 Ultra unified memory ~800 GB/s (1024-bit LPDDR5-6400). Apple-published ~800 GB/s; the full capacity is usable by the GPU - there is no VRAM ceiling because memory is unified.
Peak FLOPS
precisionfp16tflops27.2denseyesvendor_claimyessource_id—basissecondary-aggregatorbasis_detailM2 Ultra GPU ~27.2 TFLOPS FP16 class (76-core GPU variant), republished. Apple does not publish tensor-core-class figures; decode on Apple Silicon is memory-bandwidth-bound, not compute-bound, so this row is secondary to the 800 GB/s bandwidth.unit—
TDP (W)
215
Unified memory
yes
Consumer part
yes
notes
The capacity play for big dense models on used Apple Silicon. 192 GB unified at ~800 GB/s means a 70B Q4 model (~40 GB weights) reads ~40 GB/token -> ~9-11 tok/s est decode (800/40), and it can hold 70B at 8-bit plus long context or Qwen3-235B-A22B. llama.cpp canonical M-series table (discussion #4167): M2 Ultra 94.3 tok/s Q4_0 on a 7B-class model - decode scales almost perfectly with memory bandwidth. Used Mac Studio M2 Ultra 192GB ~$3,200-4,200 (Oct-2026, DRAM-shortage-inflated). Idles ~15-20 W, ~90-150 W under sustained inference, whisper-quiet. Downside: Apple RAM is soldered forever, and used pricing is inflated because new high-RAM Macs are expensive.
Notes
The capacity play for big dense models on used Apple Silicon. 192 GB unified at ~800 GB/s means a 70B Q4 model (~40 GB weights) reads ~40 GB/token -> ~9-11 tok/s est decode (800/40), and it can hold 70B at 8-bit plus long context or Qwen3-235B-A22B. llama.cpp canonical M-series table (discussion #4167): M2 Ultra 94.3 tok/s Q4_0 on a 7B-class model - decode scales almost perfectly with memory bandwidth. Used Mac Studio M2 Ultra 192GB ~$3,200-4,200 (Oct-2026, DRAM-shortage-inflated). Idles ~15-20 W, ~90-150 W under sustained inference, whisper-quiet. Downside: Apple RAM is soldered forever, and used pricing is inflated because new high-RAM Macs are expensive.
Cross-references
This record references no other records — it is a leaf. It is cited by 17.