Kvíz spojený s průvodcem · Těžko úroveň
Kvíz o využití modelových FLOPů (MFU).
Vypočítejte a interpretujte MFU při rozlišování modelových FLOP od aktivity hardwaru, špiček přesnosti a účtování závislého na architektuře.
Otázka 1 z 8
Co MFU porovnává?
Náhled každé otázky8 otázky
Odpovědi a vysvětlení se zobrazují, jak budete v kvízu.
- Co MFU porovnává?
- Measured GPU busy time divided by peak memory bandwidth
- Model parameter count divided by peak FLOPs
- Estimated useful model FLOPs per second divided by compatible peak FLOPs per second
- Kernel time divided by total optimizer-step time
- Co patří do jmenovatele při výpočtu MFU?
- Peak memory bandwidth for the same GPU
- Theoretical peak FLOPs per second for the relevant precision
- GPU activity percentage from a system monitor
- Model FLOPs per token before multiplying by throughput
- Co může způsobit, že vykazované hodnoty MFU budou nesrovnatelné?
- MFU is independent of arithmetic precision
- Models with equal parameter counts must have equal MFU
- GPU names determine an identical MFU for every workload
- Model-FLOP accounting and theoretical-peak conventions can differ
- Jak může zjednodušený odhad propustnosti tokenů tvořit MFU?
- FLOPs per token times tokens per second, divided by peak FLOPs per second
- Model FLOPs per token divided by tokens per second
- Peak FLOPs divided by model work per second
- Tokens per second divided by peak FLOPs per second
- Znamená 100procentní hlášená doba zaneprázdnění GPU 100procentní MFU?
- Yes, active kernels prove all peak arithmetic is used
- Yes, memory activity equals tensor compute
- No; activity can include memory stalls, communication, or work outside the MFU estimate
- No, because MFU measures model accuracy
- Why should published MFU values be treated as context rather than universal targets?
- MFU is independent of which accelerator is used
- Model, hardware, precision, and FLOP accounting differ
- Equal parameter counts imply the same training MFU
- Published ratios apply unchanged across precision formats
- Which pair directly shows useful training work rate and multi-GPU scaling?
- SM activity and training step time alone
- Power draw and clock frequency alone
- Training loss and validation accuracy alone
- Tokens per second and multi-GPU scaling efficiency
- What should be documented for a comparable MFU report?
- Precision, peak source, FLOP formula, workload, and measurement window
- Only the accelerator name and product-sheet peak rate
- Only achieved accuracy and total training time
- Only step time without token throughput or model-FLOP assumptions