inference/research
verifiedmedium · 0.80minorupdated 2026-10-04

Apple Silicon has better bandwidth-per-watt than discrete GPUs, but lower absolute bandwidth

[minor] hardware

Fields

id
gotchas/local-bandwidth-per-watt-apple-vs-discrete
Class
hardware
Severity
minor
Concepts
memory-bandwidthpower-efficiencybandwidth-per-wattapple-silicongddr7lpddr5x
Symptom
Apple Silicon delivers more memory bandwidth per watt than discrete GPUs, but the absolute bandwidth is lower. A Mac can run a model longer on battery, but a discrete GPU will decode faster when both are plugged in.
Root cause
Apple Silicon uses LPDR5X memory integrated on the package, which is very power-efficient but limited in absolute bandwidth (614GB/s for M5 Max, 1.2TB/s for M5 Ultra). Discrete GPUs use GDDR7 or HBM3, which deliver much higher absolute bandwidth (1792GB/s for RTX 5090) but at much higher power (575W TGP). The bandwidth-per-watt ratio favors Apple Silicon, but the absolute bandwidth favors discrete GPUs.
Workaround
For battery-powered inference, Apple Silicon is the clear winner. For plugged-in maximum performance, discrete GPUs are faster. Choose based on the use case: portable/battery (Apple) vs desktop/maximum performance (discrete).
notes
The bandwidth-per-watt comparison: M5 Max delivers 614GB/s at ~50W sustained (12.3 GB/s/W). RTX 5090 delivers 1792GB/s at ~450W sustained (4.0 GB/s/W). Apple Silicon is 3x more power-efficient per unit of bandwidth. But the RTX 5090 has 2.9x the absolute bandwidth. For a 70B model at 4-bit, the M5 Max yields ~15-25 tokens/sec on battery, while the RTX 5090 yields ~30-50 tokens/sec plugged in. The choice depends on whether portability or maximum performance matters more.

Notes

The bandwidth-per-watt comparison: M5 Max delivers 614GB/s at ~50W sustained (12.3 GB/s/W). RTX 5090 delivers 1792GB/s at ~450W sustained (4.0 GB/s/W). Apple Silicon is 3x more power-efficient per unit of bandwidth. But the RTX 5090 has 2.9x the absolute bandwidth. For a 70B model at 4-bit, the M5 Max yields ~15-25 tokens/sec on battery, while the RTX 5090 yields ~30-50 tokens/sec plugged in. The choice depends on whether portability or maximum performance matters more.

Cross-references

Affects (14)