The NVIDIA A100 is the Ampere-generation data-center GPU, 80 GB of HBM2e memory at 2 TB/s of bandwidth and 312 TFLOPS of FP16 compute. The workhorse for large-model training and inference. Available on packet.ai from $1.43/GPU-hour.
The A100 brings HBM2e memory, NVLink 3.0, and 3rd-gen Tensor Cores to data-center workloads.
The largest HBM memory of any single GPU. 70B models fit natively at FP16 without quantisation.
312 TFLOPS FP16 with sparsity up to 624 TFLOPS. The standard for large-model training.
Multi-GPU scale-up without PCIe bottleneck. Ideal for multi-node training jobs.
19.5 TFLOPS FP64 makes the A100 capable for scientific and HPC workloads too.
The standard GPU for training 7B–70B models. 80 GB HBM2e and NVLink for multi-GPU scale-up.
Run Llama 70B or similar at FP16 on a single card. No quantisation, no sharding.
Full fine-tuning or LoRA of 7B–70B models. The most memory per dollar in the lineup.
| Configuration | On-Demand | Monthly | 3 Months | 6 Months | Annually |
|---|---|---|---|---|---|
| Dedicated | Dedicated only | ||||
| 1× NVIDIA A100Most Popular | $1.43/hr | $940/mo $1.29/hr eff. | $2,820/3mo | $5,640/6mo | $11,280/yr |
| 2× NVIDIA A100Dedicated only | $2.86/hr | $1,775/mo $2.43/hr eff. | $5,325/3mo | $10,650/6mo | $21,300/yr |
| 4× NVIDIA A100Dedicated only | $5.72/hr | $3,551/mo $4.86/hr eff. | $10,653/3mo | $21,306/6mo | $42,612/yr |
The A100 is NVIDIA’s Ampere data-center GPU, 80 GB HBM2e, 312 TFLOPS FP16, NVLink 3.0. The standard for large-model training.
A100 starts at $1.43/GPU-hour dedicated. Monthly from $940/mo.
H100 has HBM3 memory (3.35 TB/s vs 2.0 TB/s) and higher FP16 throughput. A100 is cheaper and sufficient for most 70B inference and training jobs.
Full FP16 70B on a single card. For multi-GPU training, NVLink scale-up is available.
Yes. NVLink 3.0 at 600 GB/s. Available in multi-GPU cluster configurations.
Production-grade training and inference at $1.43/hr dedicated or $940/mo flat.
On-demand · hourly billing · US & EU regions