Case Studies

NVIDIA A100: Specs, Pricing and How to Reserve Your GPU VM

Written by Damanpreet Kaur Vohra | Aug 6, 2025 9:25:20 AM

What is NVIDIA A100?

The Ampere architecture powers the NVIDIA A100. It is ideal for demanding workloads like AI training, inference, HPC simulations and large-scale data processing. Hyperstack offers the following NVIDIA A100 GPUs for your workloads:

Each A100 GPU comes with 80 GB HBM2e memory, 432 third-generation Tensor Cores, FP64 performance up to 19.5 TFLOPS and AI inference up to 312 TOPS, making it highly efficient for large AI model training and scientific workloads. SXM provides 2 TB/s memory bandwidth, while PCIe versions deliver excellent compute with slightly lower interconnect performance.

Now, let’s look at how the NVIDIA A100 GPU VM is deployed on Hyperstack and what you get.

What are the Features of the NVIDIA A100 GPU?

When you deploy NVIDIA A100 GPUs on Hyperstack, you get a production-ready cloud environment with optimised storage, networking and performance for real-world AI applications

Flexible Multi-GPU Performance

Single GPUs, like the NVIDIA RTX A6000, can handle small tasks, but LLM training or multi-modal AI requires parallel GPU performance. With the A100 GPUs, you can handle such workloads with ease.

  • A100 PCIe handles small-to-mid workloads with great cost efficiency.
  • A100 PCIe NVLink and SXM allow fast inter-GPU communication, removing bottlenecks in distributed training.

You can also scale from 1 GPU to 8 GPUs in a single VM setup and handle fine-tuning to full enterprise model training without latency.

High-Speed Networking

Large-scale AI often stalls when multi-node communication is slow. Our NVIDIA A100 GPU VM offers:

  • Up to 16 Gbps networking for standard workloads and analytics via the A100 PCIe
  • Up to 350 Gbps high-speed networking for low-latency distributed training and seamless cluster scaling. It is available only for the A100 PCIe with NVLink.

Massive Memory Bandwidth and Compute Power

Training large models like GPT or Llama or running scientific simulations requires high memory throughput and raw TFLOPs. You can train bigger models and process massive datasets with fewer slowdowns. The NVIDIA A100 offers:

  • 80 GB HBM2e memory per GPU
  • Up to 2 TB/s bandwidth in NVIDIA A100 SXM
  • 432 Tensor Cores for mixed-precision AI

Ephemeral NVMe Storage for High-Speed Training

Datasets and checkpoints create I/O bottlenecks during AI training. You can benefit from the ephemeral NVMe storage in the NVIDIA A100 GPUs for temporary datasets and checkpoints

Snapshots and Bootable Volume

Reconfiguring your environment after interruptions can waste hours or even days. The NVIDIA A100 supports:

  • Snapshots to capture the entire VM state like OS, libraries and configs, allowing instant rollback or recovery during experiments.
  • Bootable Volume to ensure your OS and scripts persist across reboots, so you can pick up exactly where you left off.

NVIDIA A100 Pricing on Hyperstack

Hyperstack provides flexible GPU pricing to optimise both short-term projects and long-term enterprise workloads:

On-Demand VMs

You can access NVIDIA A100 in minutes via on-demand access. This is ideal for workloads that require immediate compute power:

  • A100 PCIe: $1.35/hour
  • A100 PCIe NVLink: $1.40/hour
  • A100 SXM: $1.60/hour
  • Billing: Pay-per-use, per-minute billing
  • Best for: Testing, experiments and short-term model training

Reserved VMs

You can reserve the same NVIDIA A100 GPU and performance for a lower price in advance but for future deployments: 

  • A100 PCIe: $0.95/hour
  • A100 PCIe NVLink: $0.98/hour
  • A100 SXM: $1.36/hour
  • Billing: Reserved pricing for fixed durations
  • Best for: Long-running LLM training, inference pipelines and research projects needing predictable costs

Spot VMs

Spot VMs for NVIDIA A100 PCIe offer cost-effective compute at just $1.08/hour, perfect for interruption-tolerant workloads like large-scale experiments, non-critical batch processing or model evaluation. However, they come with important limitations that users must plan for:

  • Spot A100 PCIe VMs do not support hibernation, high-speed networking, bootable volumes, or snapshot functionality.
  • Spot VMs can be reclaimed at any time, so use checkpointing or redundancy to avoid losing work.
  • All data stored locally is temporary and will be permanently lost upon termination.
  • Attached volumes may become corrupted or enter an error state after termination, and users are responsible for manually deleting them if needed.

Hibernate Your NVIDIA A100

You don’t need to worry about paying for idle GPU time with the NVIDIA A100 GPU VM on Hyperstack. By enabling the Hibernation feature, you can pause your VM when it’s not in use and resume it later without losing your setup. Your environment, configurations and files remain intact, so you can pick up exactly where you left off..

How to Reserve NVIDIA A100 

If your project spans weeks or months or you’re running time-sensitive model training, reserving the NVIDIA A100 on Hyperstack guarantees performance and budget control.

  • Lower pricing for long-term workloads
  • Guaranteed access during high-demand periods
  • Full usage tracking  to monitor your GPU hours

Reservation Process for NVIDIA A100

Reserving your capacity in advance is just three easy steps:

  1. Visit the Reservation Page on Hyperstack here.
  2. Complete the Form, including:
     
    • Company Name
    • Use Case (e.g., LLM training, multimodal AI, inference)
    • Number of GPUs Required (e.g., 8, 16, 32)
    • Duration (e.g., 1 month, 3 months, 6 months)

  3. Submit Your Request and our team will contact you to finalise details and ensure optimal performance for your workload.

Conclusion

The NVIDIA A100 GPU is one of the most popular choices for demanding  AI and HPC workloads. Even Meta used 16,000 A100 GPUs to train their advanced AI models like Llama and Llama 2. You can also build the next breakthrough in AI with the NVIDIA A100 GPUs and if you are still confused, which one to choose then:

  • Choose A100 PCIe for cost-effective training and inference
  • Choose A100 PCIe NVLink for distributed performance without bottlenecks
  • Choose A100 SXM for enterprise-scale LLMs, HPC, and multimodal AI

New to Hyperstack? Sign Up to Get Started Today!

Ready to Get Started?

Here are some helpful resources that will help you deploy your NVIDIA H100 SXM on Hyperstack:


FAQs

What is the NVIDIA A100 GPU?

The NVIDIA A100 is a high-performance GPU designed for AI, HPC, and data-intensive applications. It comes in three flavours on Hyperstack: A100 PCIe, A100 PCIe with NVLink and A100 SXM.

Which NVIDIA A100 GPU should I choose?

You can choose the NVIDIA A100 GPU, depending on your workloads:

  • A100 PCIe: Best for versatile, cost-effective AI training, inference, and analytics workloads.
  • A100 PCIe with NVLink: Ideal for large-scale model training with high GPU-to-GPU communication.
  • A100 SXM: Designed for extreme-scale AI and HPC with maximum memory bandwidth and NVSwitch support.

How much memory does the NVIDIA A100 GPU have?

NVIDIA A100 GPU offers 80 GB of HBM2e memory, ensuring exceptional bandwidth for large-scale AI and data analytics tasks.

What is the cost of the NVIDIA A100 GPU?

The cost of the NVIDIA A100 GPU on Hyperstack is:

  • On-Demand: Flexible pay-as-you-go starting at $1.35/hour.
  • Reserved: Lower rates for long-term commitments, starting at $0.95/hour.
  • Spot (PCIe only): $1.08/hour for interruption-tolerant workloads.

Do Spot A100 VMs have limitations?

Yes. They do not support hibernation, high-speed networking, bootable volumes, or snapshots. All data is ephemeral and can be lost if the VM is terminated.

What workloads are best suited for NVIDIA A100 GPUs?

A100 GPUs excel at large-scale AI training, high-throughput inference, scientific computing, and distributed data analytics.

How do I reserve an NVIDIA A100 GPU?

Visit the reservation page, fill in your details, submit the form and the Hyperstack team will assist you further.
View full post