Mac Studio (M3 Ultra 512GB) (as tested by Jeff Geerling): edge AI benchmarks
Best published result per workload, across model sizes — open “Show data” for like-for-like groups.
Show data — per-workload tables
object detection
No published object detection results for Mac Studio (M3 Ultra 512GB) (as tested by Jeff Geerling) yet.
We only show cited public figures — nothing is estimated to fill gaps. Requests tell us what to source or test next.
instance segmentation
No published instance segmentation results for Mac Studio (M3 Ultra 512GB) (as tested by Jeff Geerling) yet.
We only show cited public figures — nothing is estimated to fill gaps. Requests tell us what to source or test next.
pose estimation
No published pose estimation results for Mac Studio (M3 Ultra 512GB) (as tested by Jeff Geerling) yet.
We only show cited public figures — nothing is estimated to fill gaps. Requests tell us what to source or test next.
image classification
No published image classification results for Mac Studio (M3 Ultra 512GB) (as tested by Jeff Geerling) yet.
We only show cited public figures — nothing is estimated to fill gaps. Requests tell us what to source or test next.
semantic segmentation
No published semantic segmentation results for Mac Studio (M3 Ultra 512GB) (as tested by Jeff Geerling) yet.
We only show cited public figures — nothing is estimated to fill gaps. Requests tell us what to source or test next.
speech-to-text
No published speech-to-text results for Mac Studio (M3 Ultra 512GB) (as tested by Jeff Geerling) yet.
We only show cited public figures — nothing is estimated to fill gaps. Requests tell us what to source or test next.
LLM generation
| Like-for-like group | Best result | Model · precision · stack | Grade | Power | Source · date | Rank in group |
|---|---|---|---|---|---|---|
| LLM 5–14B · b=1 | 51.85 tok/s | DeepSeek R1 14b as in linked issue (Ollama default tags are 4-bit) · Ollama / llama.cpp (see linked issue) | C | 227 W measured | Jeff Geerling ai-benchmarks README (Findings tables) undated | 19 of 53 |
| LLM >40B · b=1 | 19.89 tok/s | DeepSeek R1 671b as in linked issue (Ollama default tags are 4-bit) · Ollama / llama.cpp (see linked issue) | C | 262 W measured | Jeff Geerling ai-benchmarks README (Findings tables) undated | 2 of 11 |
| LLM 15–40B · b=1 | 115 tok/s | gpt-oss 20b as in linked issue (Ollama default tags are 4-bit) · Ollama / llama.cpp (see linked issue) | C | 227 W measured | Jeff Geerling ai-benchmarks README (Findings tables) undated | 7 of 22 |
| LLM ≤4B · b=1 | 155 tok/s | Llama 3.2:3b as in linked issue (Ollama default tags are 4-bit) · Ollama / llama.cpp (see linked issue) | C | 223 W measured | Jeff Geerling ai-benchmarks README (Findings tables) undated | 17 of 64 |
How to read these numbers
Machine-readable: /find.json (see llms.txt); full tables are in “Show data”. Confidence grades: A audited MLPerf closed · B documented third-party · C community · D vendor-published · E estimate. Unknown values are shown as unknown, never estimated. Generated 2026-10-16 from public sources.