Benchmarking the Apple M1 Max
Understanding the Hardware Capabilities of Apple's flagship SOC
tlkh.dev
Summary vs stock Ryzen 5600x:
- Matrix Multiplication: 2x faster
- Numpy functions: Some wins to M1 Max, some to 5600x
- General python performance: 1.7x faster
- SpaCy NLP models: 1.25x faster
- Pandas: 1.51x faster
- GPU Matrix Multiplication (GEMM) Performance: 8 TFLOPs
- Perf/watt for training on par with RTX 3090
- M1 Max CPU gets up to 200 gb/s bandwidth vs ~40 gb/s for 5600x
- GPU used up to 330 gb/s
- CPU & GPU combined used 360 gb/s
- Idle: ~2W
- YouTube video on Google Chrome: ~3W
- MS Teams (Rosetta app), call with background blur: ~10W
- Package power seems to have a ~92W hard limit
- Fans hardly spin up from zero-RPM
- CPU seems to consume up to 40W max
- GPU seems to consume up to 60W max, usually 40-50W during benchmarks
- ANE seems to consume no more than 8W max, usually much lower (2-4W during the ResNet benchmarks)
Last edited: