Calibrated Estimator

The default estimator is the roofline bound. For higher accuracy on a specific GPU, attach a calibrated estimator (TreeEstimator): the roofline times a learned residual — one regularized LightGBM tree per operator family (GEMM, elementwise, reduction), with the bare roofline as the fallback for any uncalibrated op.

Using a calibrated model

Use the same hardware configuration you would pass to simulate or estimate_memory, including its topology. The calibration-specific parameter is:

Parameter	Meaning
calibrated_model	Path to a directory containing fitted tree models. Passing this value makes SysSim load and use TreeEstimator automatically.

from syssim import HardwareConfig

hw = HardwareConfig(
    peak_tflops_mm=1979, peak_tflops_math=989,
    peak_memory_bandwidth_GBps=3350, gpus_per_node=4,
    topology={
        "dims": ["fully_connected"],
        "size": [4],
        "bandwidth": [450],
        "latency": [12000],
    },
    calibrated_model="data/gh200",   # directory with fitted trees
)

Building a calibrated model

Profiling needs the GPU(s); calibration is CPU-only.

# 1) Profile real Megatron transformer layers over the committed shape space.
#    --num-workers N spawns N workers, one pinned per GPU.
syssim profile   --out data/gh200 --num-workers 4

# 2) Fit per-family residual trees from <data>/profile.parquet.
syssim calibrate --data data/gh200 \
    --hardware examples/configs/hardware/isambard_gh200_4gpu.yaml

There is no model-file input to profile — the shape space is the committed spec (syssim/profiling/default_spec.yaml). Preview the job list (layer configs × tensor-parallel shapes) without touching the GPU:

syssim profile --dry-run

See data/gh200/README.md in the SysSim repo for the full reproduce recipe.