Question: How much scaling behavior can be predicted before running at the target GPU scale?
I'm interested in a broader performance-engineering question related to large-scale inference/training infrastructure.
Suppose we already know the relevant workload characteristics: model architecture, parallelism, batch size, sequence characteristics, hardware configuration and network topology.
For a deployment that may scale from, say, 512 → 1,024 → 2,048 → 4,096 GPUs:
How much of the scaling curve should we realistically be able to predict before actually allocating and running at the target scale?
I'm not asking about exact throughput.
The more interesting question is whether we can predict:
the approximate scaling-efficiency band;
where the scaling curve is likely to develop a knee;
which bottleneck is likely to dominate;
and, consequently, which GPU counts are actually worth benchmarking.
The practical workflow today often seems to be:
run → measure → calibrate → extrapolate → run again.
That works, but creates an interesting asymmetry: the scale we want to evaluate may also be the scale at which assumptions calibrated at smaller scales stop holding.
I'm exploring a pre-run analysis approach where the output is deliberately not an exact prediction. Instead, it would provide:
expected efficiency band + likely scaling knee + dominant bottleneck + minimum useful validation points
The goal is not to replace benchmarking.
The goal is to make the benchmarking itself more intelligent.
From the perspective of the TensorRT-LLM/NVIDIA performance architecture:
How far do you think this can be pushed analytically before empirical calibration becomes fundamentally unavoidable?
I'm particularly interested in where the boundary lies between effects that can be modeled from workload/system parameters and effects that only become visible once the workload is actually running at scale.
Even a short maintainer perspective would be very useful.
Question: How much scaling behavior can be predicted before running at the target GPU scale?
I'm interested in a broader performance-engineering question related to large-scale inference/training infrastructure.
Suppose we already know the relevant workload characteristics: model architecture, parallelism, batch size, sequence characteristics, hardware configuration and network topology.
For a deployment that may scale from, say, 512 → 1,024 → 2,048 → 4,096 GPUs:
How much of the scaling curve should we realistically be able to predict before actually allocating and running at the target scale?
I'm not asking about exact throughput.
The more interesting question is whether we can predict:
the approximate scaling-efficiency band;
where the scaling curve is likely to develop a knee;
which bottleneck is likely to dominate;
and, consequently, which GPU counts are actually worth benchmarking.
The practical workflow today often seems to be:
run → measure → calibrate → extrapolate → run again.
That works, but creates an interesting asymmetry: the scale we want to evaluate may also be the scale at which assumptions calibrated at smaller scales stop holding.
I'm exploring a pre-run analysis approach where the output is deliberately not an exact prediction. Instead, it would provide:
expected efficiency band + likely scaling knee + dominant bottleneck + minimum useful validation points
The goal is not to replace benchmarking.
The goal is to make the benchmarking itself more intelligent.
From the perspective of the TensorRT-LLM/NVIDIA performance architecture:
How far do you think this can be pushed analytically before empirical calibration becomes fundamentally unavoidable?
I'm particularly interested in where the boundary lies between effects that can be modeled from workload/system parameters and effects that only become visible once the workload is actually running at scale.
Even a short maintainer perspective would be very useful.