Nvidia does not intend to leave the spotlight to Microsoft and Qualcomm for AI on PC. The American manufacturer intends to demonstrate the power of its RTX graphics chips with a new driver which drastically improves their performance in local AI applications.
The PC world has been in full turmoil in recent days, following recent announcements from Microsoft which inaugurate the era of “AI PCs”, under the aegis of its new Copilot+ label. These new kinds of personal computers, which integrate a multitude of services and artificial intelligence functions into the very heart of the operating system, promise to revolutionize the way we use and interact with our PCs. And whether you are an early enthusiast, a hardened skeptic or simply curious, the diversity of use cases presented by Microsoft cannot leave you indifferent.
To initiate this revolution, the Redmond firm decided to rely first of all on brand new processors, designed by the American Qualcomm, the Snapdragon X Elite and X Plus. These chips have two major characteristics: they are designed according to an Arm architecture, instead of the venerable x86, and they contain a unit specialized in operations linked to iA, the famous NPU (Neural Processing Unit or Processing Unit neuronal, in French), capable of offering a computing power of 45 TOPS, a level never reached until now for an all-in-one type processor (system-on-chip, or SoC for system-on-chip).
With its announcements and the launch of Copilot+ PCs, Microsoft is therefore propagating the idea that artificial intelligence is closely linked to Qualcomm, Arm and NPUs, at least in the Windows ecosystem. However, these are not the only players or the only technologies capable of offering high performance in terms of artificial intelligence. We must also take into account Nvidia, which has been designing and marketing, for years and well before others, particularly effective components in this area, through its RTX platform, which designates both the eponymous graphics cards well known to gamers, giant, super-powerful processors for servers and a set of software tools for development.
Nvidia R555 Game Ready: a new driver to boost local execution of AI functions
This is the new sinews of war in the era of AI PCs that is dawning. Now that the general public has discovered and adopted, at least in part, generative AI tools, it is time to move these services from the cloud to users’ machines. This is the whole purpose of the PC Copilot+ label and the Snapdragon X Series processors equipped with supercharged NPUs. But Nvidia, which has established itself as the current champion in terms of hardware components dedicated to AI, does not seem to want to stay behind in this transition and intends to return to the forefront.
As part of Microsoft’s annual developer conference, Microsoft Build, the manufacturer has just announced the release of the Nvidia Game Ready R555 driver for its RTX graphics cards, with promises of very significant gains in AI-related performance, whether for developers or users: “Large language models (LLMs) […] now work up to three times faster […] using the new NVIDIA R555 Game Ready driver.
To offer this level of performance, the new driver provides support for two software components, ONNX Runtime and DirectML, whose function is twofold: to accelerate the processing chain of iA operations by entrusting the maximum calculations to the GPU rather than the CPU (ONNX Runtime) and allow the GPU to be seen as a kind of NPU by the operating system (DirectML), so that all AI applications can address it directly, without requiring a specific implementation of RTX technologies by developers.
The objective of this driver is therefore both to drastically increase performance in terms of local AI execution and to make the computing power of the GPU much more accessible to developers and applications. To support its claims, Nvidia publishes a small graph showing the performance gains obtained on the inference (execution, as opposed to training) of three Large Language Models (LLM), with the new R555 driver compared to the previous. The progress indeed seems impressive, even if the measurements are carried out on an RTX 4090 GPU, an extremely expensive and not very common high-end card.
Keeping in mind that these figures are given by the company itself and that it will therefore be necessary to verify these results in a real situation, with different models of RTX circuits and AI applications relying on different LLMs, this communication confirms at least one thing: Nvidia does not want to be content with being present in the field of servers and training of AI models, but wants to establish itself as a key player in the execution aspect local AI services, directly on user machines.
RTX circuits have already offered, and for several years, numerous concrete applications based on AI. In video games, for example, DLSS brings together a set of scaling, image generation and graphic rendering improvement techniques based on machine learning (deep learning). Launched more recently, Nvidia ACE technology uses Large Language Models (LLM) to generate interactive dialogues in real-time with non-player characters in games. On the creative side, RTX chips accelerate many AI tools in very popular applications like Adobe Premiere, DaVinci Resolve or Blender. And the app Nvidia Broadcast offers numerous real-time processing for audio and video – and even stunning effects – on PCs equipped with an RTX chip.
There are therefore already many concrete uses of AI in the RTX ecosystem, appearing well before the general craze for generative AI, and the new Game Ready R555 driver will therefore further improve their operation. This is an interesting prospect, because Nvidia’s chips will be able to boost a wide range of artificial intelligence applications, beyond those integrated into Windows with Copilot+. It is to be hoped that publishers and developers will seize these tools to exploit the power of RTX chips in new functions based on AI, in the field of gaming, as in professional creative applications, in graphics, in 3D, in audio and video.
AI with RTX: more powerful than Copilot+ PCs?
In fact, Nvidia RTX graphics chips offer much higher levels of AI inference performance than NPU-equipped processors, whether the Qualcomm Snapdragon X Series or the upcoming Intel Lunar Lake and AMD Strix. To get an idea of this difference, the current Intel Core Ultra reaches a maximum power of 34 trillion operations per second (TOPS) and the Snapdragon X Elite and X Plus, 45 TOPS, where an RTX GPU 4050 in laptop version, the smallest model in the range, already reaches 194 TOPS. And as for the monstrous RTX 4090 in desktop version, it soars up to 1,300 TOPS!
These impressive scores are therefore well above the minimum requirement of 40 TOPS set by Microsoft to claim the title of PC Copilot+. However, this excessive power does not come without compensation and it comes at the price of much greater energy consumption than that of processors equipped with a simple NPU. And this gap in energy efficiency is widening further with the Snapdragon X Series, chips under Arm architectures whose main characteristic is precisely their very low energy consumption.
If the computing power of Nvidia’s RTX GPUs therefore allows them to tackle the heaviest and most demanding AI applications without problem, their power consumption does not make them the best candidates in the portable PC sector, in which Energy efficiency is just as important a factor as power.
An aspect that Microsoft seems to have in mind in the definition of its Copilot+ label for PC, and one of the reasons why the company reserves for the moment this name only for machines equipped with a Snapdragon X Series processor, like the ‘indicates this statement in last Monday’s announcement : “The first Copilot+ PCs will launch with Snapdragon X Elite and Snapdragon “.
However, the firm adds in the following paragraph: “New Copilot+ PC experiences will be released soon. New devices equipped with this processor combined with powerful graphics cards like NVIDIA GeForce RTX and AMD Radeon will soon be available, making Copilot+ PCs accessible to an even wider audience.” We therefore understand very clearly that the scope of PCs labeled Copilot+ is set to evolve, and that machines equipped with Nvidia RTX graphics cards could obtain the famous sesame, provided they are associated with an energy-saving central processor.
If Nvidia’s RTX GPUs therefore have ample computing power necessary to execute Copilot+ AI functions locally, obtaining the label itself escapes its control and depends entirely on Microsoft, for certainly more political considerations. and commercial rather than truly technical.
But the idea to remember is this: if you are interested in new uses of AI, you own a computer with an Nvidia RTX graphics card and Microsoft’s announcements around Copilot+ PCs have immersed you in doubt, rest assured. Your PC is far from obsolete and there is no need to run out and buy a brand new computer equipped with a processor with an NPU to use new generative AI applications locally. If media attention is currently focused on certain terms like Arm and NPU, there are other technological paths that are just as serious and relevant to propel these new uses and support the ongoing transformation of personal computing.