Workload Performance Estimator
Estimate performance on a target system from a measurement on a reference system, splitting run time into compute-bound and bandwidth-bound parts — a simple roofline-style model that never invents benchmark values.
Results
- Compute ratio
- 3.17×calculated
- Bandwidth ratio
- 1.64×calculated
- Estimated speedup
- 2.31×estimated
- Estimated performance on target
- 231.08estimated
This is an estimate derived from your measurement, not a benchmark. Validate on the target hardware.
What it calculates
- Compute and bandwidth ratios
- Estimated speedup
- Estimated performance on the target
Worked example
Using the default inputs below; change any input above to recalculate.
| Measured performance on reference | 100 |
|---|---|
| Reference peak compute | 312 TFLOPS |
| Reference memory bandwidth | 2039 GB/s |
| Target peak compute | 989 TFLOPS |
| Target memory bandwidth | 3350 GB/s |
| Share of run time that is compute-bound | 60 % |
| Compute ratio | 3.17× | calculated |
|---|---|---|
| Bandwidth ratio | 1.64× | calculated |
| Estimated speedup | 2.31× | estimated |
| Estimated performance on target | 231.08 | estimated |
Methodology
Run time is split into a compute-bound fraction that scales with peak compute and a bandwidth-bound fraction that scales with memory bandwidth: speedup = 1 ÷ (f_c ÷ R_compute + (1 − f_c) ÷ R_bandwidth).
Formulas
- Speedup
- S = 1 ÷ (f_c/R_c + (1 − f_c)/R_b)
Assumptions
- Software stack efficiency is the same on both systems.
Limitations
- Ignores interconnect, cache and software differences.