Three times cheaper, but just as efficient? This is Nvidia’s promise for its new RTX 5070.
CES 2025 was the scene of a surprising announcement from Nvidia: its new RTX 5070, sold at 649 euros, would be capable of matching the performance of the RTX 4090 marketed at 1949 euros.
To go further
Nvidia RTX 5000 vs 4000 comparison: understand everything about the new generation of graphics cards
To fully understand the scale of the challenge, let’s start by comparing the characteristics of the two cards:
Features | RTX 5070 | RTX 4090 | Differences |
Architecture | Blackwell (GB205) | Ada Lovelace | New generation |
CUDA Cores | 6 144 | 16 384 | -62,50 % |
Tensor Cores | 192 (5e gen.) | 512 (4e gen.) | -62,50 % |
RT Hearts | 48 (4e gen.) | 128 (3e gen.) | -62,50 % |
Memory | 12 Go GDDR7 | 24 Go GDDR6X | -50 % |
Bandwidth | 672 Go/s | 1 008 Go/s | -33 % |
Memory bus | 192-bit | 384-bit | -50 % |
Performance brute (FP32) | 31 TFLOPS | 83 TFLOPS | -62,70 % |
TDP | 250 W | 450 W | -44,40 % |
The RTX 5070 has 6,144 CUDA cores compared to 16,384 for the RTX 4090. A colossal difference which is found in all technical aspects: 48 SM compared to 128, 192 Tensor cores compared to 512, and a memory bandwidth of 672 GB/s compared to the 1,008 GB/s of the RTX 4090.
To the uninitiated, these figures may seem abstract. Let’s simplify: imagine the RTX 4090 as a 500 horsepower car, the RTX 5070 would be its equivalent at 185 horsepower.
These figures reveal an impressive technical gap between the two cards. Simply put, the RTX 5070 has about a third of the hardware resources of the RTX 4090. So how can Nvidia claim performance equality?
DLSS 4: Nvidia’s new weapon
The answer lies in DLSS 4, the latest evolution of Nvidia’s upscaling technology.
One of the changes in DLSS 4, Multi-Frame Generation (MFG), allows you to generate up to three additional frames between each calculated frame. An evolution compared to DLSS 3 which only generated one.
This new version also relies on a more sophisticated artificial intelligence model, moving from a convolutional neural network to a model transformeur. This change promises better image quality and reduced visual artifacts.
The latency paradox
But an essential point deserves our attention: unlike a natural increase in frameratewhich reduces latency and improves responsiveness, Multi-Frame Generation creates a paradoxical situation. Although it increases the number of images displayed, it does not improve system responsiveness. On the contrary, the AI image generation process introduces additional delay.
Nvidia attempts to compensate for this problem with a prediction system, but additional latency remains an inherent reality of this approach. For gamers sensitive to response time, especially in competition, this distinction is very important.
The concrete limits
Video memory is another significant limitation. With 12 GB of VRAM compared to 24 GB for the RTX 4090, the RTX 5070 is already in difficulty with certain recent 4K games with ray-tracing. Indiana Jones and the Great Circle, for example, already exceeds this limit.
In terms of raw performance, the gap remains considerable: 94 TFLOPS in ray-tracing for the RTX 5070 compared to 191 TFLOPS for the RTX 4090. Without DLSS, the difference in power is clearly felt.
A very conditional promise
We must therefore understand Nvidia’s announcement in context. Yes, the RTX 5070 can approach the performance of the RTX 4090, but only in certain games compatible with DLSS 4, and at the cost of slightly higher latency. For everything else — raw computing, native ray-tracing, professional applications — the power difference remains enormous.
Make no mistake: the RTX 5070 is certainly a great graphics card for its price. It even represents an impressive advance in the use of AI to improve graphics performance. But claiming that it equals the RTX 4090 is more marketing skill than technical reality.
Artificial intelligence works wonders, but it cannot yet defy the laws of physics. A lesson that Nvidia sometimes seems to forget in its marketing communication.
Find all the new features from CES 2025