Skip to content

Commit 10424ee

Browse files
committed
docs: match h3-metal comparison table and add device column
Agent-assisted-by: Codex (author) Signed-off-by: David Chen <530634352@qq.com>
1 parent 235ab66 commit 10424ee

1 file changed

Lines changed: 10 additions & 8 deletions

File tree

‎README.md‎

Lines changed: 10 additions & 8 deletions
Original file line numberDiff line numberDiff line change
@@ -94,14 +94,16 @@ Author-reported means from three requests after two warmups; not independently v
9494

9595
Author-reported results on **Apple M5 Max (40-core GPU), 128 GB, macOS 27**, at **768×448, 49 frames, seed 0**, using real weights with one engine running at a time.
9696

97-
| Engine | Steps | Cold latency (s) ↓ | Warm latency (s) ↓ |
98-
| --- | --- | --- | --- |
99-
| **h3-metal — FastH3 + block-sparse attention** | 4 | **48.23** | **18.20** |
100-
| **h3-metal — original H3** | 20 | **54.29** | **29.89** |
101-
| antirez/h3.c — default preset | 20 | 67.6 | 73.3 |
102-
| PipeNetwork/minimax-h3-mlx — 4-bit | 20 | 153.1 | 134.6 |
103-
104-
Cold includes the first clip's weight loading, text encoding and preparation; warm reuses loaded weights and prepared state. FastH3 uses a distilled 4-step checkpoint, unlike the 20-step configurations. Warm values average clips 2–4 for the 20-step rows and two A/B/B/A warm runs for FastH3.
97+
| Engine | Device | Steps | Cold | Warm | vs. MLX (cold / warm) | vs. h3.c (cold / warm) |
98+
| --- | --- | --- | --- | --- | --- | --- |
99+
| `PipeNetwork/minimax-h3-mlx` (4-bit) | Apple M5 Max (40-core GPU) | 20 | 153.1 s | 134.6 s | 1.00× / 1.00× | 0.44× / 0.54× |
100+
| `antirez/h3.c`, default preset | Apple M5 Max (40-core GPU) | 20 | 67.6 s | 73.3 s | 2.26× / 1.84× | 1.00× / 1.00× |
101+
| **h3-metal**, original H3 | Apple M5 Max (40-core GPU) | 20 | **54.29 s** | **29.89 s** | **2.82× / 4.50×** | **1.25× / 2.45×** |
102+
| **h3-metal**, FastH3 4-step + block-sparse attention | Apple M5 Max (40-core GPU) | 4 ¹ | **48.23 s** | **18.20 s** | **3.17× / 7.40×** | **1.40× / 4.03×** |
103+
104+
Cold includes the first clip's weight loading, text encoding and preparation; warm reuses loaded weights and prepared state.
105+
106+
¹ FastH3 uses a distilled 4-step checkpoint, unlike the 20-step configurations. Warm values average clips 2–4 for the 20-step rows and two A/B/B/A warm runs for FastH3.
105107

106108
[Author's report and reproduction →](https://github.com/zuiho-kai/h3-metal/blob/53b7fa5d88e2497ab427ab9c7155535f7e6e6817/README.md#performance). These external results were not rerun in InferRSI.
107109

0 commit comments

Comments
 (0)