Choose a GPU card

A GPU card and a plan are two different choices. The card is the NVIDIA accelerator, its architecture and memory; the plan is the exact vCPU count, RAM, GPU memory, CUDA version, disk space, and IOPS bundled with it, all shown together in one line on the plan list. Choose the card for the work you plan to run, then the plan size you need.

Hopper and later GPUs, the H100, H200 and B200, run only on TIR, our advanced AI/ML platform, ready-made for AI workloads. See Get started with TIR Instances.

GPU cards side by side

Card	Memory	Best for	Skip it when
NVIDIA A100 80GB	80 GB HBM2e	Multi-tenant LLM inference, training, HPC.	Your model fits in 40 GB and a cheaper plan works.
NVIDIA A100 40GB	40 GB HBM2	Mainstream LLM training and inference.	Model state exceeds 40 GB without sharding.
NVIDIA L40S	48 GB GDDR6	Cost-efficient inference, generative AI, fine-tuning.	You are training the largest models across many cards.
NVIDIA A40	48 GB GDDR6	Visualization, rendering, mid-tier inference and training.	A pure inference or training card fits your budget better.
NVIDIA A30	24 GB HBM2	Mid-range AI inference, classical ML.	Your model does not fit in 24 GB.
NVIDIA L4	24 GB GDDR6	Low-power inference, video transcoding, generative AI at scale.	You need more than 24 GB or you are training, not serving.
NVIDIA V100	32 GB HBM2	Compatibility with existing Volta-tuned pipelines.	You are starting a new project; prefer a newer card.
NVIDIA T4	16 GB GDDR6	Cost-sensitive inference and lab work.	You need modern Tensor Core throughput; prefer L4.

The live GPU tab shows the exact cards, plans, and prices for your account.

How to choose, by workload

Start with what is actually the bottleneck:

If your main concern is	Start with
Large-memory LLM inference	A100 80GB
Mid-tier training and inference	A100 40GB, or L40S
Cost-efficient generative AI inference and fine-tuning	L40S
Visualization, rendering	A40
Classical ML, smaller AI inference	A30
Low-power inference and video transcoding at scale	L4
An existing pipeline built for an older card	V100, or T4

A card has to hold the model weights plus the KV cache for your maximum context length and concurrency, not the weights alone. The KV cache is the memory a language model keeps for the text it has already read in each conversation, so it grows with longer prompts and with more conversations at once. Most out-of-memory failures at run time trace back to a card picked on price rather than memory. Sharding a model across multiple cards adds latency, so for latency-sensitive work prefer a single card whose memory comfortably fits the model over splitting it across a cheaper one.

Cards, one at a time

NVIDIA A100, 80GB and 40GB. Third-generation Tensor Cores, TF32/BF16/FP16/INT8/FP64 precisions. The 80GB variant carries more memory bandwidth than the 40GB one. Both suit LLM training and inference, HPC, and recommender systems; pick 80GB when the model state alone needs it. NVIDIA A100 80GB, NVIDIA A100 40GB.

NVIDIA L40S. 48 GB GDDR6, fourth-generation Tensor Cores. A strong cost-per-inference card with generative-AI features: stable diffusion, fine-tuning, mid-tier LLM inference, video AI. NVIDIA L40S.

NVIDIA A40. 48 GB GDDR6, built for professional visualization and mixed AI workloads: rendering, 3D, mid-tier inference. NVIDIA A40.

NVIDIA A30. 24 GB HBM2, a lower-cost datacenter card for classical ML, recommendation, and smaller-model inference. NVIDIA A30.

NVIDIA L4. 24 GB GDDR6, low-power. Built for high-volume small-model inference, video transcoding, and generative AI at a lower cost than L40S. NVIDIA L4.

NVIDIA V100 and T4. Older datacenter cards, 32 GB and 16 GB. They suit an existing pipeline already tuned for that architecture. For a new project, one of the newer cards on this page is the better fit. NVIDIA V100, NVIDIA T4.

Pricing and billing

GPU plans are billed on-demand or committed, the same two modes as any node. An on-demand GPU node is charged by the minute at its plan’s hourly price, with no minimum charge; a committed plan is paid for its whole term. The Summary panel on the create-node flow shows the live price for your selected plan, region, and billing mode. See the E2E GPU pricing page for current rates.