NVIDIA criticizes Microsoft’s idea of ​​”AI PCs”: “too basic”

NVIDIA recently presented the capabilities of its RTX GPUs for artificial intelligence, revealing that they are better than the common NPUs required by Microsoft to perform AI-related tasks. Several benchmarks were used to show how domestic GPUs outperform more modern notebooks and PCs, including Apple’s M3 Max.

According to company representatives, the 10 to 45 TOPS of performance seen in modern processors from Intel, AMD, Apple and Qualcomm are enough for just the “basic” of artificial intelligence workloads.

See more:

NVIDIA claims that tasks such as photo editing, image generation, image upscaling and AI-enhanced coding assistance on NPU-powered processors are seen to be either very basic or unable to handle – something the brand’s RTX GPUs already cover. carrying out tasks with better performance and quality.

An interesting point of the discussion is that the executives point out that RTX GPUs have 100 to more than 1,300 TOPS – depending on the model – being categorized as “premium” AI equipment, while NPUs can be seen as “basic”. She also mentions cloud computing, classified as “massive”.

NVIDIA Tests

For tests related to artificial intelligence and the difference in what is required by Microsoft, NVIDIA presented a GeForce RTX 4090 mobile GPU, an RTX 4050 mobile GPU and a Macbook Pro M3 Max with their respective benchmarks for content creation.

Stable Diffusion, Arnold, Blender, Chaos V-ray, Octane, Adobe Premier Pro Enhance Speech, DaVinci Resolve and ON1 Resize AI were also used within the test, being applied to show the difference between the hardware.

NVIDIA’s RTX 4090 managed to surpass the results presented by the M3 Max by up to seven times, in more extreme cases, and by five times in general analyses. The RTX 4050 brings twice the performance that can be seen on the Macbook.

In the LLama 2 7B int4 LLM workload, the RTX 4090 was 42% faster than the M3 Max (90% with optimizations) and the RTX 4050 performed 48% faster than the Macbook (also 90% with optimizations).

To attract attention, NVIDIA also used UL Procyon Stable Diffusion 1.5 to showcase its GPUs against the AMD Radeon RX 7900 XTX. From the GeForce RTX 4070 Super model onwards, all of them brought better results than the competition’s graphics card. The GeForce RTX 4090 showed a score 2.8 times higher.

A strong message to the industry

What NVIDIA wants with these comparisons is to show the entire industry that its GPUs are better prepared for artificial intelligence than notebooks and PCs that will arrive this year – including NPUs.

However, this demonstration is mainly aimed at Microsoft – whose definition for current “AI PCs” is to have a processor with NPU, when in fact video cards are already capable of doing all the work. Copilot, for example, does not have the option of running via GPU and this should become an even bigger trend.

According to Satya Nadella’s corporation, to be considered an “AI PC”, the computer/notebook must have an NPU alongside the CPU and GPU. This means that, for Microsoft, video cards from NVIDIA and any other manufacturer are not “qualified” for these activities.

This demonstration by Jensen Huang’s corporation not only points out that they bring better performance, but this whole idea of ​​“selling” the new “AI PCs” could be a market strategy – considering that many current computers are already ready to use intelligence artificial with what we have available.

The only point that was not mentioned by NVIDIA is the energy issue. A GeForce RTX 4090 requires two to five times more than an NPU, making possible practical use in notebooks – for example – unfeasible.

Fonte: Tom’s Hardware