Deutsch
English
In this article, you will learn more about current AI hardware. We'll explain the current hype surrounding neural processing units called NPUs and whether you really need a graphics card with AI functions like the GeForce RTX 3090. By the way, the Nvidia A100, at a price of over 20,000 Euros, is not suitable for a gaming PC ;-)
At the end of this article, you should understand whether it's time to jump on this bandwagon or hype-train, or if using artificial intelligence in the large clouds of OpenAI, Google, or Meta is still the better decision for now.
AI Hardware Looks to the Future - The GeForce 5000 Series
The demand for Nvidia's new GeForce 5000 series graphics cards is high at HI-TECH for Gamers and globally. The GeForce RTX 5080 and the GeForce RTX 5090 will be available from January 30, 2025. Of course, with such high demand, delivery will occur in batches, so we recommend acting quickly if you're interested in the latest generation of graphics cards. But how will Nvidia's new Blackwell family, as it's called, improve your Gaming PC? The answers fit perfectly into our article on AI hardware.
AI Tokens and GPUs
On January 6, 2025, Nvidia's CEO, Jensen Huang, presented the new Blackwell family at CES in Las Vegas. This was preceded by an AI-generated film that highlighted the significance of tokens.
Tokens are the smallest units of information AI works with. It's exciting to see Nvidia emphasizing the importance of these units, as their GPUs are particularly good at processing these units in parallel. Later, the focus shifted to AI support for graphics and thus gaming.
AI Can See Some Frames into the Future with Nvidia
The new generation of graphics cards offers significant performance through extensive AI training sessions. Thus, the graphics cards can already know the future coloring of pixels two to three frames before they even appear on the screen. With this technique, out of 33 million pixels (corresponding to 4K resolution), only roughly 2 million need to be rendered in the traditional way. The rest is handled by artificial intelligence.
The DLSS 4.0 software enables this advancement, which in turn is based on numerous Tensor cores - 680 in the RTX 5090.
Therefore, it's clear that anyone betting on new hardware in 2025, as installed in our gaming PCs, is also automatically betting on current AI hardware.
Some Tech Details from CES 2025 about the GeForce Blackwell Family
- 92 billion transistors
- 4,000 AI FLOPS, four times more than the last generation
- 380 Raytracing Teraflops for creating light-optimized pixels
- 125 Shader TFlops for brilliant graphics
- The RTX 5070 (available from February 2025) has a similar clock speed to the RTX 4090 at a third of the price - independent tests still need to confirm this.
.png)
What is an NPU or TPU?
Explanations of the different working methods of CPUs and GPUs can already be found on our site. The development of AI shows that chip manufacturers are quickly adapting to specialized tasks. The greatest technical challenge of our time is artificial intelligence. Hence, it's not surprising that a third type of processor has been designed for this specific area. These are the "Neural Processing Unit" (NPU) or Google's "Tensor Processing Unit" (TPU). For those who want to delve deeper into tensor calculations, tensor chips, and their differences from other processors, we can recommend this scientific paper on the subject - beware: knowledge of mathematics doesn't hurt!
We'll keep this summary brief. NPUs as well as TPUs were developed for the specific requirements of machine learning. It's about quickly training neural networks and executing trained models swiftly. For this, data parallelization is particularly important. This is where graphics cards also shine. Both types of processors can divide task packages very well and send them to various cores for simultaneous processing. CPUs, on the other hand, work more serially, meaning one task after another - of course, per core or thread.
However, NPUs have a significant advantage over GPUs. They can perform simple tasks with high repetition potential (e.g., matrix calculations) much faster.
Neural processors, besides parallel and repetitive task processing, are also designed for low energy consumption and high data throughput. The goal is to enable the fastest possible data exchange with various memories, like DRAM. Therefore, they aim for the highest possible memory bandwidth.
DeepSeek and Further AI Advances in 2025
At the end of January 2025, the Chinese company DeepSeek, with its CEO Liang Wenfeng, shook the world of artificial intelligence. Here's an easily understandable article on the topic that not only fuels the hype.
One thing the latest results from China make quite clear. The hardware requirements for AI applications are continuing to decrease. This will bring many application areas from large corporations and universities to desktops with more powerful gaming PCs. Although this might reduce sales of the expensive Nvidia H100 chipsets, the demand might shift to the GPUs found in the HI-TECH for Gamers devices:
| HI-TECH AI PCs | Component | AI Support |
|---|---|---|
|
ZOTAC Geforce RTX 4080 SUPER |
Tensor Cores 4. G | DLSS 3.5 |
|
|
ASUS TUF Gaming GeForce RTX 4090 OG OC |
1345 AI-TOPS |
|
|
INTEL Core Ultra 5 - 9 |
NPU with Intel® AI Boost |
|
|
ASUS ROG Strix Z790-F Gaming WIFI II |
AI Overclocking | Two-Way AI Noise Cancelation |
The table shows only a very small excerpt or a current snapshot. Our portfolio of AI components is expanding almost faster than some prompts can be answered ;-)
Our support team is of course happy to help you put together your desired AI hardware.
FAQs About AI Hardware in Gaming PCs
Answer: This is software to leverage Nvidia hardware capabilities to their full extent. One function is rendering games at a lower resolution. To achieve the desired quality on the monitor, machine learning then upscales the image vectors. In simple terms, communication with the game is kept as minimal as possible, and only then does artificial intelligence provide brilliant graphics in a final step.
Answer: These refer to special computational performances for AI operations. 1 TOPS stands for 1 trillion operations per second. Although the individual matrix multiplications for neural networks are very simple, the number is still remarkable.
Answer: The NPU handles computational operations that are frequently repeated when using artificial intelligence. By offloading these tasks, for example within new CPUs, the energy consumption from AI operations is significantly reduced.
Answer: Here, artificial intelligence looks at the exact load of the CPU cores and the cooling system. Then, it determines limits for performance and voltage and pushes the processors to their limit accordingly. This minimizes the risk of overloads.
Answer: Two-Way AI Noise Cancellation is exclusive software from ASUS that uses a deep learning database to remove disruptive noises like keyboard and mouse clicks as well as ambient sounds. It minimizes CPU load and hardly affects gaming performance. The function can be set separately for Input (your own voice) and Output (voices of teammates), creating a clearer and more pleasant gaming experience.
In short, disruptive ambient noises are filtered out. Such functions are especially exciting for streamers.