cedana / benchmarks · AMD MI300X · 2 GPUs
AMD MI300X cold start: ~3 minutes to 11.5 seconds
Qwen3.8-27B on vLLM 0.27 across 2 AMD MI300X GPUs. Native cold start takes 160 to 195 seconds. Cedana restores the same serving state in 11.5 seconds, 14 to 17 times faster.
160–195s
native cold start
11.5s
Cedana restore
14–17×
speedup
2
AMD MI300X GPUs
The result
About 3 minutes of cold start, back in 11.5 seconds.
Native cold start measured 160 to 195 seconds. Cedana restore took 11.5.
AMD MI300X, 2 GPUs · Qwen3.8-27B on vLLM 0.27 · speedup computed from 160–195s vs 11.5s
Inside the restore
11.5 seconds, in two parts.
CRIU restores the process in 2.7 seconds, and the GPU state comes back in 8.8 seconds. Taking the checkpoint is a separate step and is not part of the restore time: it wrote a 181 GB image in 9.4 seconds.
CheckpointWritten once, separate from the restore time
Checkpoint image181 GB
9.4s
RestoreBack to ready-to-serve, in two parts
Serving state2× AMD MI300X
11.5s
Configuration
What was measured.
- Model
- Qwen3.8-27B
- Engine
- vLLM 0.27
- Accelerator
- AMD MI300X
- GPUs
- 2
- Native cold start
- The time from launch to ready-to-serve.
- Cedana restore
- The time to bring the same serving state back from a Cedana checkpoint to ready-to-serve.
Multiply the fleet you already have.
Same hardware, same models, same serving stack. Bring one workload and we will show you the GPU-hours it gets back and the tokens it delivers.