On the other hand, the laptop RTX 5070 provides only 8GB VRAM through a 128-bit bus. It could be argued that DLSS 4’s improved AI performance reduces memory requirements, but honestly, for modern games in 2025, even at 1080p resolution, this much memory will not be enough.
What is Blackwell’s actual performance like?
Nvidia
What is the actual performance of these products? Jensen Huang didn’t go into specific performance details in his keynote, but rather delivered high-level Blackwell specifications that are more important to AI researchers than gamers. This included information such as AI TOPS, RT teraflops (TFLOPS), and “AI management processor.”
On the positive side, some of the RTX 50 series will be equipped with cutting-edge GDDR7 memory. This allows it to provide a memory bandwidth of 1.8TB/s, which is twice as much as the previous model. It also provides dual shaders for integer (INT) and floating point (FP) calculations to meet the needs of traditional game graphics, the ability to handle a mix of GPU and AI workloads as needed, and the ability to handle neural network tasks. Programmable GPU shaders are also supported.
Does it sound like Nvidia is fully invested in AI? In fact, Nvidia has been moving in this direction for several years already. Representative examples include DLSS and DLSS 3 frame generation. Jensen Huang again briefly mentioned DLSS 4 and explained that it includes functions called “neural texture compression” and “neural materials.” This is a technology that further reduces dependence on traditional GPU rendering.
According to NVIDIA, DLSS 3 can insert AI-generated frames between every frame rendered by the GPU. On the other hand, DLSS 4 can infer three full frames from just one traditional frame. AI upscaling is ultimately moving towards replacing native graphics. Jensen Huang explained that about 33 million pixels are created in four frames at 4K resolution, but with DLSS 4, the GPU renders only 2 million pixels and AI processes the rest. For more information about DLSS 4 and multiframe generation, visit NVIDIA websiteYou can check it here.
Inserting AI-generated frames also increases latency. of course NVIDIA ReflexYou can offset this to some extent by enabling , but this problem is much more noticeable when the AI is generating multiple frames. What emerged to solve this Nvidia Reflex 2Using “frame warp technology,” latency is reduced by up to 75%. The explanation of this technology is complex, so please watch the related video or refer to Nvidia’s explanation link.
If this technology works as advertised, it could bring major changes to the gaming industry.
dl-itworldkorea@foundryco.com
Source: www.itworld.co.kr
