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/s | 64-bit (4 × 16-bit) |
LPDDR5-6400 (6,400 MT/s) |
| iPhone 15 Pro (A17 Pro) |
51.2 GB/s | 64-bit (4 × 16-bit) |
LPDDR5-6400 (6,400 MT/s) |
| iPhone 16 Pro (A18 Pro) |
60.0 GB/s | 64-bit (4 × 16-bit) |
LPDDR5X-7500 (7,500 MT/s) |
| iPhone 17 Pro (A19 Pro) |
76.8 GB/s | 64-bit (4 × 16-bit) |
LPDDR5X-9600 (9,600 MT/s) |
| Huawei Pura 70 Pro+ (Kirin 9010) |
51.2 GB/s | 64-bit (4 × 16-bit) |
LPDDR5-6400 (6,400 MT/s) |
| Dell Precision 7680 (13th Gen Intel) |
89.6 GB/s | 128-bit (Dual-channel) |
DDR5-5600 |
| Desktop PC (Dual-channel DDR5) |
102.4 GB/s | 128-bit (Dual-channel) |
DDR5-6400 (6,400 MT/s) |
| MacBook Pro (M1 Pro) |
200.0 GB/s | 256-bit (4 × 64-bit) |
LPDDR5-6400 (6,400 MT/s) |
| RTX 3500 Ada (Laptop GPU) |
432.0 GB/s | 192-bit (6 × 32-bit) |
GDDR6 |
| GeForce RTX 3090 | 936.0 GB/s | 384-bit (12 × 32-bit) |
GDDR6X-19500 (19.5 Gb/s) |
| GeForce RTX 4090 | 1,008.0 GB/s | 384-bit (12 × 32-bit) |
GDDR6X-21000 (21 Gb/s) |
| GeForce RTX 5090 | 1,792.0 GB/s | 512-bit (16 × 32-bit) |
GDDR7-28000 (28 Gb/s) |
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.
Note: NVIDIA specifies the RTX 3090 with 24 GB GDDR6X over a 384-bit interface; its 19.5 Gb/s effective memory rate gives 936 GB/s. The RTX 4090 uses 24 GB GDDR6X on the same 384-bit interface at 21 Gb/s, giving 1,008 GB/s. The RTX 5090 uses 32 GB GDDR7 on a 512-bit interface at 28 Gb/s, giving 1,792 GB/s. These are theoretical peak device-memory bandwidths: (bus width ÷ 8) × effective data rate.
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 use much wider, faster GDDR memory: 432 GB/s on the RTX 3500 Ada Laptop GPU, 936 GB/s on the RTX 3090, 1,008 GB/s on the RTX 4090, and 1,792 GB/s on the RTX 5090.