3.6
Decode tok/s ↑
± 0.443 (95% CI) · sd 0.357
115
Prefill tok/s ↑ (cache-assisted)
± 6.4 (95% CI) · sd 5.15
686
TTFT ms ↓ (cache-assisted)
± 38.4 (95% CI)
274.2
Inter-token ms ↓
108.7
Peak memory GB
5
Trials
Validity: valid
The checks LemonMatrix ran while measuring; an invalid run is shown but never ranked.
- ✓ Warm-up discarded
- ✓ Thermal OK
- ✓ Exclusive run
- ✓ No model reload
- Lemonade doesn't report power draw for this instance; energy metrics omitted
- peak_memory_gb is host RAM, not VRAM -- vram_gb was unavailable from this instance
- exclusive_run verified via /api/v1/health polling (no other loaded model was busy)
Trials
Each measured trial after warm-up, from the run's trial sidecar. The summary above is computed from these.
| # | ttft ms | decode tokens per sec | prompt tokens | output tokens | host memory gb | decode tokens per sec |
|---|---|---|---|---|---|---|
| 1 | 686.5 | 3.924 | 79 | 158 | 107.2 | |
| 2 | 698.8 | 4.123 | 79 | 133 | 107.8 | |
| 3 | 649.2 | 3.314 | 79 | 153 | 108 | |
| 4 | 666 | 3.374 | 79 | 160 | 108.3 | |
| 5 | 729.6 | 3.503 | 79 | 173 | 108.7 |
Same model, other runs
Every run of Qwen3.8-27B-GGUF-UD-Q4_K_XL · UD-Q4_K_XL on the board, ranked by decode tok/s. This run is highlighted.
| Profile | Device | Engine | Backend | decode tok/s ↑ | Valid | |
|---|---|---|---|---|---|---|
| AMD-Instinct-MI350X | gfx950 GPU | dgpu | llamacpp-rocm | 100.01 | valid | |
| Radeon-RX-7900-XTX | AMD Radeon RX 7900 XTX | dgpu | llamacpp-vulkan | 20.48 | valid | |
| Radeon-RX-7900-XTX-WSL | AMD Radeon RX 7900 XTX (gfx1100) | dgpu | llamacpp-rocm | 17.10 | valid | |
| Radeon-RX-7900-XTX | AMD Radeon RX 7900 XTX | dgpu | llamacpp-rocm | 14.19 | valid | |
| Radeon-RX-7900-XTX | AMD Radeon RX 7900 XTX | dgpu | llamacpp-vulkan | 9.83 | invalid | |
| NVIDIA-L4 | Intel(R) Xeon(R) Silver 4314 CPU @ 2.40GHz | cpu | llamacpp-cpu | 3.65 | valid | |
| Radeon-RX-7900-XTX | AMD Radeon RX 7900 XTX | dgpu | llamacpp-rocm | 2.50 | invalid |
Model and configuration
| model.name | Qwen3.8-27B-GGUF-UD-Q4_K_XL |
| model.class | dense |
| model.quantization | UD-Q4_K_XL |
| model.context_length | 4096 |
| config.compute_engine | cpu |
| config.backend | llamacpp-cpu |
| config.os | linux |
| config.power_state | plugged |
Environment
| cpu | Intel(R) Xeon(R) Silver 4314 CPU @ 2.40GHz |
| device_model | Intel(R) Xeon(R) Silver 4314 CPU @ 2.40GHz |
| dgpu | 7x NVIDIA L4 |
| dgpu_family | 7x sm_89 |
| driver_version | 535.309.01 |
| lemonade_version | 11.6.0 |
| memory_gb | 251.5 |
| os_version | Linux-6.8.0-117-generic (Ubuntu 22.04) |
| backend_version | b10375 |
All metrics
| prefill.tokens_per_sec | 115.2 |
| prefill.stddev | 5.152 |
| prefill.p95 | 121.1 |
| prefill.ci95_half_width | 6.396 |
| decode.tokens_per_sec | 3.647 |
| decode.inter_token_latency_ms | 274.2 |
| decode.stddev | 0.3571 |
| decode.p95 | 4.084 |
| decode.ci95_half_width | 0.4433 |
| ttft_ms | 686 |
| trial_count | 5 |
| peak_memory_gb | 108.7 |
| prompt_sha256 | 1c43b07d6b0d7624785966ffe2ddc076aa76b5fcf9357882ac1165f0073f43ee |
| ttft_ms_stddev | 30.9 |
| ttft_ms_ci95_half_width | 38.36 |