📊 TimesFM 2.5 200M Benchmark Report
Generated: 2026-07-04T03:28:43.867Z · Git: e1c65b8d · Node: v22.23.1
· ONNX Runtime 1.27.0
💻 System
CPU
AMD EPYC 9V74 80-Core Processor × 4
🧠 Model
⚡ Inference Latency — Node.js vs WASM
| Context | Batch |
Node Avg (ms) | Node P99 (ms) |
WASM Avg (ms) | WASM P99 (ms) |
WASM/Node |
| 128 |
1 |
64.2 |
83.3 |
330.4 |
345.8 |
5.1× |
| 128 |
2 |
122.5 |
124.3 |
631.0 |
665.5 |
5.2× |
| 128 |
4 |
248.9 |
264.8 |
1295.7 |
1521.6 |
5.2× |
| 128 |
8 |
579.6 |
924.0 |
— |
— |
— |
| 256 |
1 |
118.1 |
172.4 |
390.6 |
393.5 |
3.3× |
| 256 |
2 |
271.7 |
485.2 |
783.5 |
797.9 |
2.9× |
| 256 |
4 |
292.5 |
533.9 |
1559.6 |
1564.5 |
5.3× |
| 256 |
8 |
490.3 |
495.9 |
— |
— |
— |
| 512 |
1 |
62.1 |
68.2 |
389.8 |
391.6 |
6.3× |
| 512 |
2 |
122.4 |
125.8 |
780.1 |
782.1 |
6.4× |
| 512 |
4 |
244.3 |
245.5 |
1578.7 |
1602.9 |
6.5× |
| 512 |
8 |
507.9 |
597.6 |
— |
— |
— |
🖥️ Node.js Inference Latency
| Context | Batch | Patches |
Avg (ms) | P50 (ms) | P99 (ms) |
Throughput (seq/s) | Cold Start (ms) |
| 128 |
1 |
4 |
64.2 |
61.3 |
83.3 |
15.6 |
63.8 |
| 128 |
2 |
4 |
122.5 |
122.5 |
124.3 |
8.2 |
61.1 |
| 128 |
4 |
4 |
248.9 |
245.6 |
264.8 |
4 |
61.6 |
| 128 |
8 |
4 |
579.6 |
517.3 |
924 |
1.7 |
60.9 |
| 256 |
1 |
8 |
118.1 |
114.7 |
172.4 |
8.5 |
107.1 |
| 256 |
2 |
8 |
271.7 |
228.3 |
485.2 |
3.7 |
108.6 |
| 256 |
4 |
8 |
292.5 |
246.6 |
533.9 |
3.4 |
222.6 |
| 256 |
8 |
8 |
490.3 |
489.8 |
495.9 |
2 |
61.2 |
| 512 |
1 |
16 |
62.1 |
61.1 |
68.2 |
16.1 |
63.8 |
| 512 |
2 |
16 |
122.4 |
122 |
125.8 |
8.2 |
63.9 |
| 512 |
4 |
16 |
244.3 |
244.5 |
245.5 |
4.1 |
61 |
| 512 |
8 |
16 |
507.9 |
490.2 |
597.6 |
2 |
62.3 |
🌐 WASM Inference Latency
| Context | Batch |
Avg (ms) | P50 (ms) | P99 (ms) |
Throughput (seq/s) | Cold Start (ms) |
| 128 |
1 |
330.4 |
329.4 |
345.8 |
3 |
382.8 |
| 128 |
2 |
631 |
624.4 |
665.5 |
1.6 |
336.2 |
| 128 |
4 |
1295.7 |
1275.7 |
1521.6 |
0.8 |
311.8 |
| 256 |
1 |
390.6 |
390.1 |
393.5 |
2.6 |
392.4 |
| 256 |
2 |
783.5 |
780.4 |
797.9 |
1.3 |
393.2 |
| 256 |
4 |
1559.6 |
1559.9 |
1564.5 |
0.6 |
392.9 |
| 512 |
1 |
389.8 |
389.8 |
391.6 |
2.6 |
390.7 |
| 512 |
2 |
780.1 |
779.9 |
782.1 |
1.3 |
389.7 |
| 512 |
4 |
1578.7 |
1578.3 |
1602.9 |
0.6 |
389.7 |
🎯 Prediction Accuracy
Full TimesFM pipeline with RevIN normalization, flip invariance, continuous quantile head
| Metric | Value |
|---|
| Naive MAE (no-change) | 87.7782 |
| Constant Mean MAE | 60.6676 |
| TimesFM MAE | 6.9042 |
| TimesFM RMSE | 8.4396 |
| Scaled MAE vs Naive | 0.0787 ✅ Better |
| Improvement vs Naive | 92.1% |
| Improvement vs Const Mean | 88.6% |
💾 Memory Footprint (Node.js)
| Metric | Value |
|---|
| RSS | 901 MB |
| Heap Used | 10 MB |
| Heap Total | 12.5 MB |