GPU Model Guides

By Architecture

Hopper, Blackwell, Ada Lovelace, Ampere, CDNA — every GPU lineage available in cloud.

In this category

CDNA 3 powers the MI300X and MI325X — AMD's answer to H100/H200 with bigger HBM capacity per package.

CDNA 4 (MI350X, MI355X) is AMD's next-gen frontier compute architecture with 288 GB HBM3e per GPU.

Ada Lovelace covers L4, L40, L40S, RTX 4090, RTX 6000 Ada, and the broader RTX 40-series. The price/perf sweet spot for inference and fine-tuning.

Ampere — A100, A40, A30, A10, A2, RTX A6000, RTX 30-series — is the most widely available GPU family in cloud, often at the…

Blackwell is NVIDIA's current-generation AI architecture. Every Blackwell GPU available across cloud providers.

Blackwell Ultra is NVIDIA's frontier-class refresh with the highest VRAM and memory bandwidth in production today.

Hopper-class GPUs (H100, H200, GH200) are still the workhorse of large-model training and inference. Compare every Hopper GPU in the cloud.

Pascal-era GPUs (Tesla P4, GTX 1080) are the cheapest end of the cloud GPU market — useful for legacy inference and experimentation.

Turing GPUs (Tesla T4) remain a budget pick for inference and video workloads.

Volta (Tesla V100) is legacy but still cheap and capable for ML research, fine-tuning, and inference.