GPU Model Guides

By VRAM

Filter cloud GPUs by VRAM capacity, from 12 GB up to 288 GB+.

In this category

Every cloud GPU with at least 12 GB of VRAM — the floor for running modern small-to-mid LLMs, image-gen, and most fine-tuning jobs.

141 GB+ VRAM — H200-class and above. The minimum for serving Llama-3.1 405B or DeepSeek-V3 on a single node.

Cloud GPUs with 16 GB+ VRAM — comfortable for SDXL inference, 7B-13B model fine-tuning, and most production inference workloads.

192 GB+ VRAM — Blackwell-class and MI300X. Maximum on-device capacity per GPU for trillion-parameter regime workloads.

GPUs with 24 GB+ VRAM unlock 13B-30B model inference, larger batch sizes, and longer context windows.

256 GB+ VRAM — frontier AI training territory. MI325X, MI350X, MI355X, B300, GB200.

288 GB+ VRAM — the absolute top tier of single-GPU memory capacity available today.

GPUs with 32 GB+ VRAM — the entry point for serious training and 30B+ model fine-tuning without sharding.

48 GB+ VRAM is the sweet spot for fine-tuning 30B-70B models on a single GPU and for production multi-tenant inference.

64 GB+ VRAM — covers premium professional workloads and the larger data-center class GPUs.

80 GB+ VRAM is the default for frontier AI training (A100 80GB, H100, H200, B200, MI300X). Compare every option side-by-side.

96 GB+ VRAM — for 70B+ model training without sharding and for multi-GPU inference of the largest open models.