| VRAM | System RAM | |
|---|---|---|
| Where | On the graphics card | On the motherboard |
| Used by | GPU | CPU (and GPU via a slower path) |
| Typical bandwidth | Hundreds to ~1,800 GB/s on current GPUs | Far lower on dual-channel desktop DDR5 |
| Upgradeable? | No: fixed per card | Usually yes |
For AI, a model that fits in VRAM runs much faster than one split between VRAM and RAM, because generation speed is limited by memory bandwidth. Unified memory (Apple silicon, AMD Ryzen AI Max, NVIDIA DGX Spark) is one large pool shared by CPU and GPU; see unified memory vs VRAM.
GPU memory bandwidth, cited
| GPU | VRAM | Memory bandwidth | Launch MSRP (US) | Source |
|---|---|---|---|---|
| GeForce RTX 5090 | 32 GB | 1792 GB/s | $1,999 | NVIDIA: GeForce RTX 5090 specs; NVIDIA Newsroom (Jan 2025): RTX 5090/5080/5070 Ti/5070 pricing; Wikipedia: GeForce RTX 50 series (memory bandwidth table) |
| GeForce RTX 5080 | 16 GB | 960 GB/s | $999 | NVIDIA: GeForce RTX 5080 specs; NVIDIA Newsroom (Jan 2025): RTX 5090/5080/5070 Ti/5070 pricing; Wikipedia: GeForce RTX 50 series (memory bandwidth table) |
| GeForce RTX 5070 | 12 GB | 672 GB/s | $549 | NVIDIA: GeForce RTX 5070 family specs; NVIDIA Newsroom (Jan 2025): RTX 5090/5080/5070 Ti/5070 pricing; Wikipedia: GeForce RTX 50 series (memory bandwidth table) |
| GeForce RTX 5060 | 8 GB | 448 GB/s | $299 | NVIDIA: GeForce RTX 5060 family specs; NVIDIA Newsroom (Apr 2025): RTX 5060 family pricing; Wikipedia: GeForce RTX 50 series (memory bandwidth table) |
VRAM and bus width from NVIDIA spec pages; bandwidth from Wikipedia series tables (not listed on NVIDIA pages); MSRPs from NVIDIA press releases. Current street prices: TBD.
How much system RAM to buy: how much RAM do I need.