Cascade
Trace an ML workload once on a real GPU. Cascade replays it on simulated hardware, across as many GPUs as you want.
how it works
Finding out how a model runs on hardware you don’t have means renting that hardware, if you can get it at all. Testing on 64 GPUs means paying for 64 GPUs.
Cascade starts from one real run. Our profiler hooks a live vLLM server with Accel-Sim’s NVBit tracer and streams every kernel trace off the GPU while it is still running.
The trace lands in the Cascade app, a Sail sandbox. One launch forks that sandbox into many simulators. Each fork already has the trace on disk and runs Accel-Sim 2.0 against its own GPU config.
Every simulator reports back live with kernel progress, simulated cycles and IPC. When real hardware timing exists, Cascade lines up simulated against real, kernel by kernel.
Next, the simulated GPUs act as one system. Cascade splits an inference workload across them, so the whole fleet serves it together.
links
team
Built at the Neo Hackathon.