GPUs with 128–∞ GB VRAM for Running LLMs
Browse 10 GPUs with 128–∞ GB VRAM compatible with running LLM models locally. Compare VRAM, memory bandwidth, and AI performance. Top picks and budget guide
10 GPUs
| GPU | Price |
|---|---|
| H200 SXMPro141 GBHBM3e · 4,800 GB/s · 700 WNVIDIA · Hopper · 2024 | – |
| B200Pro192 GBHBM3e · 8,000 GB/s · 1000 WNVIDIA · Blackwell · 2024 | – |
| H200 NVLPro141 GBHBM3e · 4,800 GB/s · 600 WNVIDIA · Hopper · 2024 | – |
| B300Pro288 GBHBM3e · 8,000 GB/s · 1400 WNVIDIA · Blackwell Ultra · 2025 | – |
| Instinct MI300XPro192 GBHBM3 · 5,300 GB/s · 750 WAMD · CDNA 3 · 2023 | – |
| Instinct MI325XPro256 GBHBM3e · 6,000 GB/s · 1000 WAMD · CDNA 3 · 2024 | – |
| Instinct MI350XPro288 GBHBM3e · 8,000 GB/s · 1000 WAMD · CDNA 4 · 2025 | – |
| Instinct MI355XPro288 GBHBM3e · 8,000 GB/s · 1400 WAMD · CDNA 4 · 2025 | – |
| GH200 Grace Hopper SuperchipPro144 GBHBM3e · 4,900 GB/s · 1000 WNVIDIA · Hopper (Grace Hopper) · 2024 | – |
| Instinct MI250XPro128 GBHBM2e · 3,277 GB/s · 560 WAMD · CDNA 2 · 2021 | – |
Which GPU Do You Need for AI?
The amount of VRAM is the most important specification for running LLMs locally. Most 7B parameter models require 4–8 GB of VRAM at common quantization levels, while 70B models need 24–48 GB. Memory bandwidth determines how fast the model generates tokens — faster bandwidth means faster responses.