NVIDIA
Ampere A100 SXM4
312 FP16 · 80GB HBM2e · 400W
61Score
TrainingFine-tuning
Specifications
ArchitectureAmpere
Memory80GB HBM2e
Memory Bandwidth2,039 GB/s
FP16 TFLOPS312
FP8 TFLOPS0
BF16 TFLOPS624
INT8 TOPS1,248
TDP400W
InterconnectNVLink 3.0 (600 GB/s) (600 GB/s)
EcosystemCUDA
GenerationPrevious
Est. Price$12,000
Memory
Generation
HBM2e
Capacity
80 GB
Bandwidth
2.0 TB/s
2,039 GB/s
Bus Width
5,120-bit
HBM Vendor(s)SK Hynix / Samsung
Recommended Configuration
8× A100 in DGX A100
Training Intelligence
◆CUDA
◆PyTorch
◆TensorFlow
◆JAX
◆DeepSpeed
Training Time Estimates
LLaMA 70B(70B)
~28 days64 GPUs
GPT-3 175B(175B)
~34 days1024 GPUs
Stable Diffusion XL(3.5B)
~2.5 days8 GPUs
Cloud cost: $3.67/hr
Rent this GPU in the Cloud
regularly tracked, last cycle October 2026 · rate updated outside cycle
No hardware purchase required — rent by the hour from these providers.
Thunder Compute$1.09/GPU/hr
Rent on Thunder →Hyperstack$1.35/GPU/hr
Rent on Hyperstack →RunPod$1.59/GPU/hr
Rent on RunPod →Hyperstack$1.60/GPU/hr
Rent on Hyperstack →Crusoe$2.00/GPU/hr
Rent on Crusoe →Crusoe$2.30/GPU/hr
Rent on Crusoe →FluidCore$2.40/GPU/hr
Rent on FluidCore →FluidCore$2.49/GPU/hr
Rent on FluidCore →CoreWeave$2.70/GPU/hr
Rent on CoreWeave →Lambda$2.79/GPU/hr
Rent on Lambda →Azure$3.56/GPU/hr
Launch on Azure →GCP$5.07/GPU/hr
Launch on GCP →