A40
nvidia · ampere · 48 GB GDDR6 · 300W TDP
VRAM
48 GB
BF16 TFLOPS
37.4
Bandwidth
696 GB/s
From
$0.42/hr
Spec Sheet
Pricing by Provider
| Provider | On-Demand | Reserved | Spot | Badge |
|---|---|---|---|---|
| tensordock | $0.59/hr | - | $0.42/hr | Cheapest |
| vast_ai | $0.65/hr | - | $0.44/hr | |
| runpod | $0.89/hr | - | $0.65/hr | |
| lambda | $0.79/hr | - | - |
Compatible Models (267)
Single GPU (198 models)
Multi-GPU (69 models)
Training Capabilities
Estimated GPU count for full fine-tuning (AdamW, BF16) and QLoRA
| Model Size | Full Fine-Tune | QLoRA |
|---|---|---|
| 7B model | 3 GPUs | 1 GPU |
| 13B model | 6 GPUs | 1 GPU |
| 70B model | 28 GPUs | 1 GPU |
Energy Efficiency
Estimated tokens/second per Watt for popular models
Similar GPUs
| GPU | VRAM | BF16 TFLOPS | BW (GB/s) | From |
|---|---|---|---|---|
| RTX A6000 | 48 GB | 38.7 | 768 | $0.49/hr |
| A100 40GB SXM | 40 GB | 312 | 1555 | $0.85/hr |
| A100 40GB PCIe | 40 GB | 312 | 1555 | $0.69/hr |
| A16 | 64 GB | 16.8 | 232 | $0.72/hr |
| A10G | 24 GB | 35 | 600 | $0.30/hr |
Methodology Note
Performance estimates for the A40are based on InferenceBench's roofline performance model with CUDA kernel-level optimization including FlashAttention v2 and PagedAttention. Memory calculations account for model weights (48 GB GDDR6 available), KV-cache allocation, and activation memory. Throughput predictions use the A40's rated 696 GB/s memory bandwidth and 37.4 BF16 TFLOPS compute capacity as roofline ceilings, with empirical correction factors per GPU architecture (ampere). See our full methodology.
Frequently Asked Questions
How many AI models can run on A40?
The A40 can run 267 AI models from our database within a single node. Compatible models range across various parameter sizes depending on the quantization precision (BF16, FP8, INT4). Smaller models fit on a single GPU while larger models may require multi-GPU setups up to 8x A40.
What is the A40 inference throughput?
The A40 delivers 37.4 BF16 TFLOPS and 37.4 FP8 TFLOPS with 696 GB/s memory bandwidth. Actual inference throughput (tokens/sec) depends on the model size, precision, and batch size. Use our calculator for model-specific throughput estimates.
How much does A40 cost per hour?
The A40 is available starting from $0.42/hour via tensordock. Prices vary by provider and pricing tier (on-demand, reserved, spot). Compare pricing across all providers in the table above.