GPU catalogue

Blackwell Ultra flagship for frontier training and long-context inference. Largest HBM per GPU in the catalogue.

VRAM

288GB

FP16

2.5 PFLOPS

Rate

$7.50/hr

Rack-scale Grace Blackwell system billed per GPU. For workloads that need 72 GPUs in one NVLink domain.

VRAM

186GB

FP16

2.25 PFLOPS

Rate

$6.90/hr

Blackwell datacenter GPU with 180GB HBM3e. Suited for training and serving the largest open-weight models on a single node.

VRAM

180GB

FP16

2.25 PFLOPS

Rate

$5.50/hr

AMD's CDNA 4 flagship with 288GB HBM3e and FP4/FP6 support. ROCm-first stack for large-model serving.

VRAM

288GB

FP16

2.3 PFLOPS

Rate

$4.50/hr

Hopper with 141GB HBM3e and 4.8TB/s bandwidth. The sweet spot for serving 70B–400B models with long context.

VRAM

141GB

FP16

989 TFLOPS

Rate

$3.60/hr

CDNA 3 with 256GB HBM3e. Memory-heavy inference for models that do not fit on 141GB.

VRAM

256GB

FP16

1.31 PFLOPS

Rate

$3.20/hr

PCIe H100 with 94GB and an NVLink bridge between pairs. Built for LLM inference in standard servers.

VRAM

94GB

FP16

835 TFLOPS

Rate

$2.80/hr

The workhorse of modern AI. SXM form factor with NVLink for multi-GPU training and high-throughput inference.

VRAM

80GB

FP16

989 TFLOPS

Rate

$2.69/hr

192GB HBM3 per GPU — runs a 70B model in BF16 on a single card. Strong vLLM and SGLang support.

VRAM

192GB

FP16

1.31 PFLOPS

Rate

$2.40/hr

PCIe H100 for single-GPU fine-tuning and inference where NVLink is not required.

VRAM

80GB

FP16

756 TFLOPS

Rate

$2.20/hr

Intel's training and inference accelerator with 128GB HBM2e and built-in Ethernet scale-out. Runs PyTorch via Habana SynapseAI.

VRAM

128GB

FP16

1.83 PFLOPS

Rate

$2.00/hr

Hopper GPU fused to a 72-core Grace CPU with 480GB LPDDR5X. Huge unified memory for inference that spills past HBM.

VRAM

96GB

FP16

989 TFLOPS

Rate

$1.85/hr

Ampere datacenter GPU. Still the best value for training mid-size models and serving 7B–70B models.

VRAM

80GB

FP16

312 TFLOPS

Rate

$1.60/hr

Previous-gen CDNA 2 accelerator with 128GB. Plentiful and cheap on HPC-oriented clouds.

VRAM

128GB

FP16

383 TFLOPS

Rate

$1.50/hr

PCIe A100 for standard servers. Same compute as SXM with lower interconnect bandwidth.

VRAM

80GB

FP16

312 TFLOPS

Rate

$1.40/hr

The original A100. 40GB is enough for most fine-tuning runs under 13B parameters.

VRAM

40GB

FP16

312 TFLOPS

Rate

$1.10/hr

Ada datacenter GPU for inference, rendering and video. FP8 support and 48GB make it a strong single-GPU server.

VRAM

48GB

FP16

362 TFLOPS

Rate

$0.90/hr

Visual-computing Ada GPU. Good for diffusion models, rendering and moderate LLM inference.

VRAM

48GB

FP16

181 TFLOPS

Rate

$0.80/hr

Ampere visual-computing GPU with 48GB. Cheap capacity for inference that needs memory more than speed.

VRAM

48GB

FP16

150 TFLOPS

Rate

$0.40/hr

Low-power Ada inference card. Ideal for small models, video transcoding and embeddings at scale.

VRAM

24GB

FP16

121 TFLOPS

Rate

$0.40/hr

Legacy Volta GPU. Cheap tensor cores for classic deep learning and small fine-tunes.

VRAM

16GB

FP16

125 TFLOPS

Rate

$0.20/hr

70W inference card. The cheapest CUDA GPU on most clouds — fine for embeddings and tiny models.

VRAM

16GB

FP16

65 TFLOPS

Rate

$0.15/hr

Deskside Grace Blackwell Ultra with 784GB coherent memory. A single box that fits trillion-parameter inference.

VRAM

784GB

FP16

2.5 PFLOPS

Rate

$9.00/hr

NVIDIA's desktop AI supercomputer: Grace Blackwell GB10 with 128GB unified memory. Hosted by the community — runs 200B-parameter models locally.

VRAM

128GB

FP16

125 TFLOPS

Rate

$0.90/hr