Post

Memory bandwidth comparison across devices

Memory bandwidth comparison across devices

A comparison of theoretical peak memory bandwidth, bus widths, and memory standards across mobile SoCs, laptop workstations, desktop PCs, and discrete laptop GPUs:

Device /
Processor
Peak
bandwidth
Memory bus
width
Memory
standard
iPhone 14 Pro
(A16 Bionic)
51.2 GB/s64-bit
(4 × 16-bit)
LPDDR5-6400
(6,400 MT/s)
iPhone 15 Pro
(A17 Pro)
51.2 GB/s64-bit
(4 × 16-bit)
LPDDR5-6400
(6,400 MT/s)
iPhone 16 Pro
(A18 Pro)
60.0 GB/s64-bit
(4 × 16-bit)
LPDDR5X-7500
(7,500 MT/s)
iPhone 17 Pro
(A19 Pro)
76.8 GB/s64-bit
(4 × 16-bit)
LPDDR5X-9600
(9,600 MT/s)
Huawei Pura 70 Pro+
(Kirin 9010)
51.2 GB/s64-bit
(4 × 16-bit)
LPDDR5-6400
(6,400 MT/s)
Dell Precision 7680
(13th Gen Intel)
89.6 GB/s128-bit
(Dual-channel)
DDR5-5600
Desktop PC
(Dual-channel DDR5)
102.4 GB/s128-bit
(Dual-channel)
DDR5-6400
(6,400 MT/s)
MacBook Pro
(M1 Pro)
200.0 GB/s256-bit
(4 × 64-bit)
LPDDR5-6400
(6,400 MT/s)
RTX 3500 Ada
(Laptop GPU)
432.0 GB/s192-bit
(6 × 32-bit)
GDDR6

Note: The NVIDIA RTX 3500 Ada Laptop GPU typically comes configured with 12 GB of GDDR6 VRAM.
Note: The Huawei Pura 70 Pro+ memory bandwidth figure (51.2 GB/s) is referenced from Wikipedia’s HiSilicon processor database (calculated from 3,200 MHz / LPDDR5-6400 over a 64-bit bus). However, Wikipedia does not currently include a direct citation for this specific metric, and web searches have not yet yielded an authoritative primary source.

Key observations

  • Smartphone SoCs: Moving from LPDDR5-6400 (51.2 GB/s on A16/A17 Pro) to LPDDR5X-7500 (60.0 GB/s on A18 Pro) and LPDDR5X-9600 (76.8 GB/s on A19 Pro) steadily increases mobile memory bandwidth without widening the 64-bit bus. The Kirin 9010 on the Huawei Pura 70 Pro+ maintains 51.2 GB/s over a standard 64-bit quad-channel LPDDR5-6400 bus.
  • Desktop DDR5: Standard dual-channel DDR5-6400 on a desktop platform achieves 102.4 GB/s of memory bandwidth across its 128-bit bus.
  • Apple Silicon unified memory: The M1 Pro achieves 200 GB/s by employing a wide 256-bit bus, feeding both CPU and GPU from a shared high-bandwidth pool.
  • Discrete GPUs vs. host RAM: While standard dual-channel DDR5 delivers ~76–102 GB/s, discrete GPUs like the RTX 3500 Ada utilize high-speed GDDR6 to reach 432 GB/s over a 192-bit bus.