In this article you can read about,

• what characterizes NVIDIA's A100 graphics processing unit (GPU),
• which application scenarios are possible for companies and
• how you can use the computing power flexibly as required in the cloud.

What was once a favorite for gamers has long been used by researchers, engineers and companies. They all appreciate the enormous computing power of modern graphics processing units (GPUs). NVIDIA is currently at the forefront of this development with the A100 GPU, which enables a wide range of application scenarios.

The market for graphics processing units (GPUs) has experienced impressive momentum in recent years. The performance of graphics cards has increased dramatically and has created impressive experiences for PC users with graphics-intensive applications. But through their architecture, GPUs also bring the ability to serve high-performance requirements. The splitting of large workloads and concurrent processing are an important enabler for high performance computing in science and research. Last but not least, graphics cards also play a key role in the training and inference of artificial intelligence (AI). GPU manufacturers have designed specific graphics cards for use in data centers for such high-end purposes.

It is therefore no coincidence that NVIDIA positions itself as the “world's leading provider of AI computing”. With its products, NVIDIA is constantly setting new benchmarks for GPU performance. The current price/performance benchmark for AI training is the A100 GPU.

Even competitors such as the MI250 from MosaicML or chips manufactured by AMD cannot currently hold a candle to the A100. The A100 GPU is based on the Ampere architecture and has around 20 times the computing power of the previous generation. In combination with the highest memory bandwidth in the world (over two terabytes per second), it delivers unprecedented computing power.

The A100 is not just designed for one specific use case: it trains AI models, calculates complex simulations or analyzes large amounts of data. But the true added value of the A100 lies in its flexibility and ability to handle multiple workloads simultaneously. Multi-instance GPU technology makes it possible to divide the A100 into up to seven independent instances, allowing several small jobs to be handled simultaneously.

The A100 GPUs support all modern high-performance requirements for simulation, modeling and AI in various industries. In many cases, they even make deployment scenarios possible or shorten their cycle – for example from weeks to days – thanks to the reduced computing time. This allows companies more frequent insights and better control options. Some examples:

• Energy:
  Smart grids are becoming a reality thanks to the A100: energy consumption forecasts and real-time analysis of grid performance are becoming faster and more accurate. The real-time analysis of data from wind turbines optimizes their operation, but the analysis of seismic data to identify new energy deposits is also accelerated.
• Production and maintenance:
  In predictive maintenance concepts, forecasting algorithms evaluate the condition of machines and identify maintenance requirements at an early stage. With the A100 GPU, faster and more detailed evaluations are also possible here, enabling optimal production planning and minimizing downtimes.
• Health:
  The A100 enables faster training of deep learning algorithms. The use (inference) of deep learning services is also accelerated: doctors and researchers receive faster analysis of medical data for studies and prognoses. Individual diagnostics also benefit from the increase in speed.

It's no secret that quality comes at a price. Buying an A100 is an expensive undertaking: List prices are around €20,000 – if you can get one.

The cloud offers an elegant way to use an A100 on demand – whenever it is needed – and thus significantly reduces or spreads the costs. On the Open Telekom Cloud, the use of an A100 as a p3.2xlarge.8 flavor with 80GB HBM2 main memory currently (October 2023) costs €3.57 per hour. That's well over 5,000 operating hours compared to a purchase. And costs are only incurred – in the cloud sense – when the GPU is used.

If a GPU is heavily used and utilized in the long term, a purchase can be considered – if you want to put in the management effort. But for users who only need high-performance resources sporadically or temporarily, using them from the cloud makes more business sense. In addition, the GPU is available at all times and the “delivery time” is significantly shorter.

At the same time, this demonstrates the “democratic principle” of the cloud: it also gives smaller companies access to high-end computing capacities. A comprehensive analysis over the weekend costs just under €200 – a hundredth of the purchase price and an amount that is affordable even for the smallest companies. Incidentally, the Open Telekom Cloud also offers discounted reserved packages for long-term use.

So buy or rent? In the latter case, companies use a flexible alternative that meets their needs and avoids the high acquisition costs. Because one thing is certain: the A100 does not mark the end of GPU technology – competitors are not sleeping and are bringing dynamism to the market. Wherever GPU performance will develop in the future: the cloud gives companies access to the latest GPU generations – flexibly and in line with their needs.