deploybase

A100 on Paperspace: Pricing, Specs & How to Rent

Deploybase · February 5, 2025 · GPU Pricing

Contents

A100 GPU on Paperspace: Availability & Pricing

Paperspace offers A100 GPUs on its Cloud GPU platform. As of March 2026, availability varies by region. Good option for medium to large-scale ML workloads.

The NVIDIA A100 GPU delivers 80 GB of GPU memory (HBM2e), making it suitable for transformer model training, fine-tuning, and inference serving. Paperspace integrates A100 access with pre-configured deep learning templates, Jupyter notebooks, and containerized environments.

Standard A100 configs include persistent storage, managed networking, integrated billing. Production customers get dedicated support and custom infrastructure.

A100 Technical Specifications

The A100 has 6,912 CUDA cores across 108 Streaming Multiprocessors. Peak performance: 312 TFLOPS (BF16/FP16 tensor), 19.5 TFLOPS (FP32). The A100 does not support native FP8 (FP8 was introduced with Hopper/H100).

Memory capacity reaches 80 GB of HBM2e (or 40 GB variant) with 1.935 TB/s bandwidth (PCIe 80GB). This memory configuration handles most production transformer models including Llama 2 70B and Mistral 8x7B at reasonable batch sizes.

The A100 features third-generation NVIDIA NVLINK with 600 GB/sec interconnect speed between paired GPUs. This enables efficient distributed training across multiple A100s when Paperspace configurations include multi-GPU setups.

TensorCore units provide specialized acceleration for matrix operations. The GPU executes structured sparsity patterns for reduced compute requirements in certain workloads.

Power consumption is 400W TDP (SXM) or 300W (PCIe 80GB), significantly lower than H100s (700W). Paperspace allocates power budgets accordingly across shared infrastructure.

A100 on Paperspace: Rental Cost Analysis

Paperspace A100 pricing depends on machine type and region. Standard on-demand A100 instances cost approximately $3.18 per hour for an A100 80GB as of March 2026.

A typical A100 + vCPU + RAM + storage machine is billed at the published per-hour rate through Paperspace.

Monthly calculations show:

  • On-demand: $3.18/hour = $76.32/day = $2,321/month (730 hours)
  • Three-month discount: 10% reduction to approximately $2,089/month
  • Annual commitment: 20% reduction to approximately $1,857/month

Paperspace Gradient provides a managed training platform layered on top of GPU infrastructure, adding overhead for managed orchestration.

Spot instances on Paperspace offer discounts but include interruption risk.

How to Rent A100 on Paperspace

Account Creation:

  • Sign up at paperspace.com
  • Add payment method and billing details
  • Verify email address
  • Set up SSH keys for terminal access

Launching an A100 Instance:

  • Navigate to Console > Machines > Create
  • Select "GPU" machine type
  • Choose A100 from GPU options
  • Select region (US-East, West-Coast, or EU options available)
  • Choose OS: Linux (Ubuntu 20.04 or 22.04) or Windows Server
  • Configure vCPU and RAM allocation
  • Set storage size (minimum 50 GB recommended)
  • Launch instance

Accessing The A100:

  • SSH into machine using provided IP address
  • Or use Paperspace web terminal in browser
  • Pre-installed CUDA toolkit 11.8 or 12.x depending on OS
  • NVIDIA drivers already configured

Running Training Workloads:

  • Clone model repositories via git
  • Install Python dependencies with pip
  • Upload datasets to persistent /storage volume
  • Launch training scripts using standard PyTorch or TensorFlow commands
  • Monitor GPU usage with nvidia-smi

A100 Performance Across Key Workloads

Large language model training shows strong throughput on A100. A 7-billion parameter model achieves 95,000 tokens/second training throughput with batch size 16.

Fine-tuning demonstrates efficient memory utilization. A 13-billion parameter Llama model fine-tunes in 5 hours using QLoRA adapters with rank 64 and batch size 8.

Inference serving with text generation models reaches 35 tokens/second for Llama 2 70B with bfloat16 quantization.

Image generation using Stable Diffusion XL produces high-quality outputs at 8 seconds per 1024x1024 image.

Multi-GPU setups scale near-linearly. Two A100 GPUs achieve 1.95x throughput compared to single GPU due to NVLINK interconnect overhead.

A100 vs. Alternative GPUs on Paperspace

The A100 balances cost and capability effectively. For comparison, H100 specifications deliver higher single-GPU throughput but at 3-4x cost premium.

Paperspace's L40S GPUs cost 15-20% less than A100 and offer 48 GB GDDR6 memory. L40S suits inference and batch processing well; however, A100's HBM2e memory provides superior bandwidth for large training runs.

The RTX 4090 costs 40% less than A100 but struggles with transformer models exceeding 20-30 billion parameters due to 24 GB memory limitation.

For teams prioritizing cost, Paperspace spot A100 instances at 50% discount provide excellent training economics if workload interruption tolerance exists.

Integration with Paperspace Ecosystem

Paperspace Gradient provides CI/CD pipeline management for reproducible training runs. Teams commit model code to GitHub, and Gradient handles GPU allocation, dependency installation, and result logging automatically.

Tensorboard integration visualizes training metrics directly in Paperspace console without additional setup.

Paperspace Notebooks enable Jupyter-based interactive development. Notebooks persist across A100 instance power cycles when data stored on /storage volume.

Integration with cloud storage services: Paperspace machines can mount Amazon S3, Google Cloud Storage, or Azure Blob using standard CLI tools.

Model hosting on Paperspace enables serving trained models as API endpoints. Teams deploy inference containers on Paperspace machines with automatic scaling based on request load.

When A100 Makes Economic Sense

A100 on Paperspace suits teams with recurring training needs spanning 40+ hours monthly. At $3.18/hour, monthly spend reaches $1,272 for 400 hours of continuous training.

Single projects lasting 1-2 weeks benefit from on-demand pricing without annual commitments. The no-contract model lowers barrier to experimentation.

Spot instances become attractive for non-critical workloads tolerating occasional interruption. Research prototyping or batch inference jobs recover cost 50% through spot pricing.

Teams requiring geographic proximity benefit from Paperspace's EU regions. GDPR-regulated workloads can remain in European infrastructure without transatlantic data transfer.

FAQ

Q: What operating system should I select for A100 on Paperspace?

Linux (Ubuntu 22.04) provides faster package installation and broader compatibility with deep learning frameworks. Windows Server works but carries licensing costs.

Q: Can I run multiple models simultaneously on a single A100?

Yes, using NVIDIA MPS (Multi-Process Service) or vLLM for inference. Training a single large model typically consumes full GPU capacity.

Q: Is data stored on the persistent volume encrypted?

Paperspace encrypts data at rest on managed storage. Additional encryption via LUKS can be configured for enhanced security if needed.

Q: How long does A100 provisioning take?

Most instances launch within 2-3 minutes. Regional availability impacts provisioning speed; US regions typically provision faster than EU.

Q: Can I schedule A100 instances to pause and resume?

Yes, Paperspace Console provides scheduling for automatic stop/start on defined times. This reduces costs for dev/test workloads used during business hours only.

Q: What's the maximum instance size on Paperspace?

Standard on-demand A100 instances come single-GPU. Multi-GPU A100 setups require contacting Paperspace production sales.

A100 Specs Guide - Complete technical specifications

GPU Pricing Guide - Compare all providers

Paperspace GPU Pricing - Detailed rate information

RunPod GPU Pricing - Alternative provider comparison

LLM Fine-Tuning Guide - Training methodology

Sources

  • NVIDIA A100 Tensor GPU Technical Specifications
  • Paperspace GPU Cloud Documentation
  • Paperspace Pricing and Machine Specifications
  • NVIDIA CUDA Toolkit Documentation
  • Deep Learning Framework Performance Benchmarks