huawei-cloud-ascend-op-mfu-calculator

Calculate MFU (Machine FLOP Utilization) for operators like matmul/GEMM/FlashAttention on Ascend NPU, providing clear formulas and derivation process Use when the user wants to: (1) calculate MFU for matrix operations, (2) analyze operator performance efficiency, (3) bottleneck diagnosis from Ascend profiling data, (4) optimize operator implementation Triggers include: "MFU", "machine flop utilization", "operator FLOPs", "matmul performance", "GEMM efficiency", "Ascend MFU", "算子MFU", "算力利用率", "矩阵乘效率", "GEMM性能", "FlashAttention性能", "瓶颈诊断", "bottleneck"

Huawei Cloud Ascend Operator MFU Calculator

Overview

This skill calculates MFU (Machine FLOP Utilization) for operators like matmul/GEMM/FlashAttention on Ascend NPU, providing clear formulas and derivation process.

Architecture: Input Validation → FLOPs Calculation → Achieved TFLOPs/s → MFU Calculation → Result Analysis

Related Skills:

• Profiling data analysis skills may provide operator performance data (execution time and memory traffic) used by this skill's MFU calculation and bottleneck diagnosis.

Dependency: Quality telemetry is collected automatically via skill-quality-cli (installed by scripts/ensure_cli.sh if absent).

Prerequisites

1. Python 3.8+ installed
2. Basic understanding of FLOPs calculation concepts

• skill-quality-cli — ensured by bash scripts/ensure_cli.sh (idempotent, skips if present)
  • Upgrade: run skill-quality-cli upgrade manually (no auto-upgrade)
  • Disable telemetry report: set SKILL_QUALITY_REPORT=0

All hcloud commands in this skill are wrapped with skill-quality-cli run for telemetry. No manual action required.

Usage Scenarios

Typical Problem Scenarios:

• Evaluating how well an operator utilizes Ascend NPU compute power
• Comparing performance of different operator implementations
• Identifying optimization opportunities for matrix operations

Typical User Utterances:

• "Calculate MFU for my GEMM operator"
• "What's the machine FLOP utilization for FlashAttention?"
• "Analyze my matmul operator performance efficiency"

Workflow

1. Input Collection: Gather operator parameters (matrix dimensions, data types, execution time)
2. FLOPs Calculation: Compute theoretical FLOPs for the operation
3. Achieved Performance: Calculate achieved TFLOPs/s from execution time
4. MFU Calculation: Apply formula MFU = Achieved FLOPs / Peak FLOPs
5. Result Analysis: Provide interpretation and optimization suggestions

MFU Calculation Formula

MFU = (Achieved FLOPs / Peak FLOPs) × 100%

Where:

• Achieved FLOPs = Operation FLOPs / Execution Time
• Peak FLOPs = Hardware-specific peak performance. Use the per-model FP16/BF16 peak values below — do NOT use a single generic value like 256 TFLOPs.

Ascend Model	FP16/BF16 Peak (TFLOPs/s)
910B1	≈ 378.88
910B2	≈ 353.89
910B3	≈ 294.91
910B4	≈ 270

Reference Documents

Document	Description
Ascend 910B Series Technical Specifications	Official Ascend 910B series product specifications
MFU Calculation Methodology	Detailed MFU calculation formulas and examples
FlashAttention Technical Paper	Original FlashAttention research paper

Enhanced Features

Intelligent Bottleneck Diagnoser

A rule-based bottleneck diagnoser (scripts/bottleneck_diagnoser.py) that analyzes Ascend profiling CSV data (e.g. OpBasicInfo.csv from msprof) and automatically classifies each operator's bottleneck:

• memory-bound — achieved memory bandwidth utilization is close to the hardware peak
• compute-bound — AI Core (compute) utils are the limiting resource
• communication-bound — collective communication operators (HCCL AllReduce/AllGather, etc.) dominate
• operator-fallback — the operator ran on AI CPU / CPU instead of an AI Core kernel

It computes per-operator achieved TFLOPS, memory bandwidth and compute utilization, then returns a ranked report (priority P0/P1/P2) with concrete optimization recommendations.

bash

skill-quality-cli run --skill-name huawei-cloud-ascend-op-mfu-calculator -- python3 scripts/bottleneck_diagnoser.py --csv OpBasicInfo.csv --peak-tflops 294.91 --peak-memory-bw 1200

Column matching is tolerant of the exact msprof header names (e.g. Task Duration(us), AI Core Time(us), Memory Read Bytes, Memory Write Bytes). If a column is absent, the related signal is skipped and classification falls back to the remaining signals.

Core Commands

The skill is a pure calculation/diagnosis skill (no Huawei Cloud CLI mutation): everything is computed locally with Python 3.8+.

Command	Purpose
MFU calculation	Compute FLOPs, achieved TFLOPs/s and MFU with the formulas in MFU Calculation Methodology; use per-model peak TFLOPs from the table above and confirm the execution-time unit (ms vs us)
python3 scripts/bottleneck_diagnoser.py --csv <profiling.csv> [--peak-tflops <TFLOPs>] [--peak-memory-bw <GB/s>]	Auto-classify operator bottlenecks from Ascend profiling CSV

Parameter Confirmation

Parameter	Description	Required
operator	Operator type (matmul/flash_attention/gemm, etc.)	Yes
flops	Theoretical FLOPs of the operator	Yes
time_ms	Operator execution time (milliseconds)	Yes
time_us	Operator execution time (microseconds) — alternative to time_ms	No
peak_tflops	Hardware peak computing power (TFLOPS); use the per-model table above	Yes
device	NPU device type (910B1/910B2/910B3/910B4/910, etc.)	No

Time Unit Handling (ms / us)

MFU is sensitive to the execution-time unit — mixing ms and us changes the result by ~1000×. Follow these rules:

• time_ms is the canonical input (milliseconds).
• If the user provides the duration from an Ascend profiler (e.g. profiling CSV columns Task Duration(us)), that value is in microseconds (us) — convert it before calculating:
  • time_ms = time_us / 1000
  • Achieved TFLOPs/s = flops / (time_us × 1e-6) / 1e12 (equivalently flops / (time_ms × 1e-3) / 1e12)
• Alternatively pass time_us as the input parameter and let the calculation convert internally.
• Always ask the user to confirm the unit when it is ambiguous, and state the unit in the final MFU result.