Publication Date
2 Nov 2026
Reading Time
11 min
Writer
vMind Team
A GPU (Graphics Processing Unit) is a processor built to run the same mathematical operation on thousands of pieces of data at the same time. Its first job was drawing images on a screen. Today, most of the work of training and running AI models happens on GPUs.
When a team starts an AI project, the first question is usually what hardware it will need. This guide covers what a GPU is, why it works differently from a CPU and which types exist. After that, we compare data center GPU models for AI workloads and look at the question we hear most often: how does the H100 differ from the H200? vMind offers these cards in its Istanbul data centers, and you can also browse all of our AI solutions.
İçindekiler
What Is a GPU?
A GPU is a type of processor that splits a job across a large number of small cores and runs them in parallel. A CPU has somewhere between a few and a few dozen powerful cores. In the GPUs used in data centers, the count runs into the thousands: the SXM version of the NVIDIA H100 has 16,896 CUDA cores. Each of these cores is weaker than a single CPU core on its own. When they all work at once, though, repetitive calculations such as matrix multiplication finish far sooner.
The product that made the term “GPU” popular was the GeForce 256, which NVIDIA released in 1999. Back then, a graphics card’s job was to draw the 3D scenes in video games. When CUDA came out in 2007, developers could start programming GPUs for calculations that had nothing to do with graphics. The real turning point for AI came in 2012: an image recognition model called AlexNet was trained on two GTX 580 gaming cards and won that year’s ImageNet competition by a wide margin. From then on, training deep learning models on GPUs became the norm.
What Does GPU Stand For?
GPU is an abbreviation of Graphics Processing Unit. It is also often called simply a graphics processor. Many people who search for “what does GPU mean” are actually thinking of a graphics card, so it is worth separating the two ideas.
Is a GPU the Same as a Graphics Card?
Not exactly. The GPU is the processor chip that sits on a graphics card. The graphics card is the whole board: along with that chip, it carries the memory (VRAM), power circuitry, cooler and display outputs. Most laptops and phones have no separate card at all; the GPU is built into the same chip as the main processor. And many data center GPUs have no display output whatsoever, because they are made purely for computation.
What Is a GPU Used For?
In a home or office computer, the GPU keeps games, video editing software and browser animations running smoothly. In businesses, the range of uses is much wider. Every task below repeats the same calculation millions of times, and that is exactly where the GPU’s advantage comes from:
- Training large language models and fine-tuning them on company data (model training infrastructure)
- Serving a trained model so it answers users in production, in other words inference
- Medical image analysis and camera-based quality control on production lines
- 3D rendering, video processing and engineering simulations
- Financial risk modeling and analysis of large datasets
How Does a GPU Work?
Picture a 12-megapixel photo. It contains 12 million pixels, and you need to apply the same filter to each one. A CPU shares this job among its handful of cores and processes the pixels largely one after another. A GPU spreads the pixels across thousands of cores and processes them simultaneously. Training a neural network is mathematically similar: large matrices are multiplied together, the weights are updated based on the result, and this loop repeats millions of times.
Cards designed for AI also include Tensor cores dedicated to this work. Tensor cores perform matrix multiplication at lower precisions such as FP16 or FP8, and at much higher speed.
Core count alone does not decide how a card actually performs. Two more figures matter: memory capacity and memory bandwidth. If a model does not fit in the card’s memory, it has to be split across several cards. If bandwidth falls short, the cores sit idle waiting for data. Most of the difference between the H100 and the H200 comes down to these two figures.
The Difference Between a CPU and a GPU
CPUs and GPUs are often presented as rivals, but inside a GPU server they work together. The CPU prepares the data, sends it to the GPU and collects the results, while the GPU handles the heavy computation. This division of labor is why vMind GPU servers allocate 16 vCPUs and 128 GB of system memory per card. For general-purpose workloads such as websites, databases or ERP, CPU-based Compute servers are sufficient and more economical.
| Core design | A small number of powerful cores. Usually between 4 and 24 in desktop processors; server processors can exceed 100. | Thousands of simple cores. The NVIDIA H100 SXM has 16,896 CUDA cores. |
| How it works | Sequential, branching, decision-heavy tasks | Running the same operation in parallel across large amounts of data |
| Memory | System RAM; high capacity, relatively low bandwidth | HBM or GDDR memory on the card; very high bandwidth (4.8 TB/s on the H200) |
| Good at | Operating systems, databases, web servers, business applications | Model training, inference, rendering, simulation |
| Struggles with | Large matrix calculations | Heavily branching code that follows sequential logic |
| Role in the system | Manages the system and hands work to the GPU | Carries out the heavy computation sent by the CPU |
GPU Types: iGPU, dGPU, Data Center GPU and Cloud GPU
How many kinds of GPU are there? The answer depends on where the GPU sits and where it gets its memory from.
| Integrated GPU (iGPU) | On the same chip as the processor; laptops, office PCs, phones | Shares system RAM | Office work, watching video, light gaming |
| Discrete GPU (dGPU) | On a separate graphics card plugged into the motherboard | Its own VRAM (GDDR6, GDDR7) | Gaming, design, video editing, small model experiments |
| Data center GPU | Inside servers, usually in groups of 8 cards connected via NVLink | HBM (high-bandwidth memory) | Model training, inference, HPC |
| Cloud GPU | In the provider’s data center, accessed over the network | The memory of the allocated card | GPU rental and GPU as a Service |
When companies talk about GPUs for AI projects, they almost always mean the cards in the third row. A single H100 SXM card draws up to 700 watts, and these cards are mostly installed in servers of 8. An office server room cannot handle that load; it takes a data center environment. The cloud GPU in the fourth row is the way to reach the same power without owning these servers.
Which GPU for AI?
Data Center GPU Models The first two figures to check when choosing a card are memory and bandwidth. Below is a summary of the models offered on vMind’s GPU infrastructure; the values match the technical specifications published on NVIDIA’s official product pages.
| NVIDIA H100 | Hopper | 80 GB HBM3 | 3.35 TB/s | Large model training, distributed training |
| NVIDIA H200 | Hopper | 141 GB HBM3e | 4.8 TB/s | Long context, large model inference |
| NVIDIA L40S | Ada Lovelace | 48 GB GDDR6 | 0.86 TB/s | Inference, image processing, rendering |
| NVIDIA A100 | Ampere | 80 GB HBM2e | 2.0 TB/s | Fine-tuning, mid-scale training |
What Is the Difference Between the NVIDIA H100 and H200?
The H100 and H200 share the same Hopper architecture and the same compute cores; the difference between them lies almost entirely in memory. In place of the H100’s 80 GB of HBM3, the H200 carries 141 GB of HBM3e, and bandwidth rises from 3.35 TB/s to 4.8 TB/s. So the H200 does not compute any faster. It holds a larger model on a single card and moves data to the cores more quickly.
| Architecture | Hopper | Hopper |
| Memory | 80 GB HBM3 | 141 GB HBM3e |
| Memory bandwidth | 3.35 TB/s | 4.8 TB/s |
| Compute performance (FP8, FP16) | Same | Same |
| Power consumption (SXM) | Up to 700 watts | Up to 700 watts |
| Where it shines | Compute-heavy training, models that fit in memory | Memory-heavy inference, long context, large models |
An example makes the practical difference easy to see. A 70-billion-parameter model takes up about 70 GB at FP8 precision. It fits on a single H100, but that leaves only around 10 GB for the working memory (KV cache) needed when handling long texts. On an H200, the same model runs with roughly 70 GB to spare, which means longer context and more concurrent users on the same card. At FP16 precision, the same model grows to about 140 GB and has to be split across at least two H100s.
That is why the decision depends on whether the workload is bound by compute or by memory. If the model fits comfortably in memory and the work is mostly training, the H100 delivers the same compute power at a lower cost. If the model is large, the context is long, or you are building an inference service that answers many requests in production, the H200’s extra memory can cut the number of cards you need and bring the total cost down.
A Practical Guide to Choosing a Card
- If you are training a model on a multi-card cluster and the model fits in 80 GB of memory: NVIDIA H100 rental
- If the model does not fit in the H100’s memory, or you are building an inference service that works with long texts: NVIDIA H200 rental
- Image processing, rendering and serving mid-sized models: NVIDIA L40S
- Fine-tuning an open-source model on company data: the NVIDIA A100 is enough for most projects
If you are not sure which card suits you, our AI consulting team can size your workload with you.
Getting GPU Power Without Buying GPUs
Buying the hardware is not the only way to have data center-class GPUs. Companies can use these cards from the cloud under two models. With GPU rental, a card is allocated to you and your own team runs the software and operations on it. With GPU as a Service, setup, driver updates, 24/7 monitoring and capacity planning also move over to the provider.
At vMind, both models run on the same hardware in Istanbul-based, ISO 27001-certified data centers; data never leaves Turkey and invoices are in Turkish lira. Our GPU rental or purchase? page, where we compare the total cost of buying and renting line by line, can help guide your decision.
Frequently Asked Questions
What is a GPU, in short?
A GPU (graphics processing unit) is a processor that runs the same operation on thousands of pieces of data at the same time. It was developed to draw graphics; today, most of the work of training and running AI models happens on GPUs.
Are a GPU and a graphics card the same thing?
No. The GPU is the processor chip on the graphics card. The graphics card is the whole board, which carries the memory, power circuitry, cooler and display outputs alongside the GPU.
Isn’t a CPU enough for AI?
A CPU can do the job for experimenting with small models. But for work that demands heavy parallel computation, such as model training and inference in production, the time difference between a CPU and a GPU is measured in hours, sometimes days. For general-purpose workloads, CPU-based servers are still the most economical option.
Should I choose the H100 or the H200?
If the model fits comfortably in 80 GB of memory and the work is mostly training, the H100 is more economical. If the model is large, you work with long context, or you are running a busy inference service, the H200’s 141 GB of memory lets you do the same job with fewer cards.
You can find answers to other questions about AI and GPU infrastructure in our Knowledge Center.
Choose the Right GPU with vMind
vMind reviews your workload, model size and budget, then recommends the best setup in terms of card model, number of cards and service model. We offer H100, H200, L40S and A100 cards in our Istanbul data centers, either with hourly usage or as a managed service.
For a pilot project, you can get started quickly with hourly usage. If you are planning an enterprise-scale cluster, you can pin down your requirements with a solutions architect through our workload assessment. For pricing tailored to your card model, quantity and duration, contact the vMind team.
