NVIDIA GeForce RTX 5070 Founders Edition Review – When AI has to help out on the finish line

1 - Introduction and details of the GB205-300 2 - Test system and equipment 3 - Gaming: Full-HD 1920x1080 Pixels (Rasterization Only) 4 - Gaming: WQHD 2560x1440 Pixels (Rasterization Only) 5 - Gaming: Ultra-HD 3840x2160 Pixels (Rasterization Only) 6 - Gaming: WQHD 2560x1440 Pixels, Supersampling, RT & FG 7 - Gaming: Ultra-HD 3840x2160 Pixels, Supersampling, RT & FG 8 - DLSS4 and MFG: Cyberpunk 2077 in detail 9 - DLSS4 and MFG: Alan Wake 2 in detail 10 - PCIe 5 problems, power consumption in theory and practice 11 - Load peaks native vs. DLSS4, PSU recommendation 12 - Cooler, temperatures, thermography, noise 13 - Summary and conclusion

Yes, what have we here? With the GeForce RTX 5070 FE, NVIDIA has now launched a new graphics card on the market that leaves me somewhat perplexed in many respects. I’m deliberately putting this up front, because you have to read today’s review carefully and also get involved in the subtleties, which I will highlight more visually this time. No, I’m certainly not going to write a rant, even if you have to ask yourself what NVIDIA is trying to achieve with this card. Because one thing is clear: the development is already very visibly reaching its physical limits.

Important preliminary remarks and a paradigm shift

The continuous increase in raster performance has been the central driver for advances in graphics hardware for decades. Modern architectural improvements, higher memory bandwidths and more efficient manufacturing processes have made it possible to steadily increase the performance of GPUs. However, this path now seems to have reached its physical and economic limits, at least that’s the impression I got with the GeForce RTX 5070. Despite the new architecture and minimal increase in efficiency, it barely achieves higher performance in classic raster graphics calculations than its direct predecessor, the RTX 4070 Super.

The main reason for this is that the scaling of the computing units can no longer be continued on the same scale as in the past. The sudden increase in performance by simply adding more shader units or by increasing the clock frequency is becoming increasingly inefficient. NVIDIA and other manufacturers are therefore increasingly relying on alternative technologies, especially artificial intelligence. Features such as DLSS (Deep Learning Super Sampling) or AI-supported image and physics calculations are intended to increase the perceived performance without having to significantly increase the raw performance of the GPU. The GeForce RTX 5070 mainly benefits from such new AI functions, but whether this strategy will be enough to convince the market remains to be seen.

One potential stumbling block is, of course, the broad acceptance of such technologies. While AI-supported upscaling methods such as DLSS are already convincing in many scenarios, the question remains to what extent users are prepared to do without traditional rendering methods. Success also depends largely on software support. If developers and gamers do not perceive the subjective added value of AI functions as sufficient progress, the RTX 5070 could appear less attractive than expected despite the technical innovations. In addition, the 32-bit CUDA libraries are no longer supported, making PhysX a prominent victim in older games. Why NVIDIA does not implement a wrapper here, however, is beyond my understanding.

Ultimately, this development shows a paradigm shift: pure hardware performance enhancement is increasingly being replaced by algorithmic optimization. The next few years will show whether this approach is sufficient to meet the high expectations of new GPU generations. And now back to the card.

The recommended retail price (RRP) for the NVIDIA GeForce RTX 5070 in Germany is 600 euros. But who is the card actually aimed at? I see it in the area of Full HD and maximum WQHD resolutions, because the 12 GB graphics memory of the GeForce RTX 5070 could increasingly become a bottleneck in modern games, especially in Ultra HD resolution (3840 × 2160 pixels). A particularly striking example of this is The Last of Us Part I on the PC. This game is known for its high VRAM requirements, especially at maximum texture details and in high resolutions. Even on an RTX 4070 Super with 12 GB VRAM, the game regularly reaches its memory limit in Ultra HD. This not only leads to possible performance drops, but also to a visible loss of display quality when the driver starts to play tricks.

Although the RTX 5070 brings with it a new architecture that can offer efficiency gains, the memory size itself remains unchanged. Given the increasing VRAM requirements of many new titles – not just The Last of Us Part I, but also games like Hogwarts Legacy or Alan Wake 2 – this could call into question the future-proofing of the card. NVIDIA often argues that modern memory compression techniques and AI-supported upscaling (such as DLSS) can help to reduce memory requirements. In practice, however, it has been shown that these measures do not completely eliminate the problem. Especially for users who want to play in native Ultra HD without major compromises, the GeForce RTX 5070 may therefore not be the best choice if alternative models with 16 GB or more memory are available. Of course, this limits the target group somewhat. So let’s agree on a maximum of WQHD or proper AI support for supersampling, shall we?

Unfortunately, I also have to mention that I had to switch the card to PCIe 4 this time, as I had to deal with sporadic boot loops as well as black screens, which made exact benchmarking very difficult. Switching the PCIe generation down to 4 worked of course, but it is a pure workaround that does not solve the actual problem of the Founders Edition. However, I assume that the card is not fast enough for this disadvantage to be noticeable in the results.

I already mentioned that NVIDIA has focused less on a drastic increase in raster performance with the RTX 5070, but instead has created a more efficient GPU through architectural improvements, larger cache structures and optimized AI functions. This is also reflected in the support of DLSS 4, which, in combination with the improved Tensor cores, should enable even more realistic rendering. The Blackwell architecture thus brings significant advances in the way rendering and image enhancement are linked, even if the pure raw performance in classic raster graphics applications does not go beyond what was already possible with the previous generation.

The GB205-300 GPU in detail

The GeForce RTX 5070 is based on NVIDIA’s new Blackwell architecture and uses the GB205-300 graphics processor, which is manufactured using the advanced 4 nm process. This chip comprises around 31.1 billion transistors and covers an area of 263 mm², making it significantly more compact than the GB202 of the high-end models. Despite its moderate size, the chip offers high efficiency and performance, which is achieved through an optimized shader architecture, new cache structures and increased RT and tensor performance.

A total of 6,144 CUDA cores are organized inside the GB205 processor, which are distributed across several Graphics Processing Clusters (GPCs). Each of these GPCs contains several Texture Processing Clusters (TPCs), which in turn consist of Streaming Multiprocessors (SMs). The SMs represent the basic computing unit of the GPU and, in addition to the classic FP32 and INT32 cores, also contain specialized hardware for ray tracing and AI calculations. With the Blackwell architecture, NVIDIA relies on the fourth generation of RT cores and the fifth generation of Tensor cores, which should significantly accelerate both ray tracing performance and AI-supported image enhancement.

A key innovation is the improved cache hierarchy. The GB205 chip has a 48 MB L2 cache, which optimizes memory accesses and increases the effective bandwidth of the graphics memory. This is particularly relevant as the Geforce RTX 5070 is equipped with a 192-bit memory interface, which works with 12 GB GDDR7 memory, but offers less raw bandwidth compared to high-end cards with wider interfaces. Thanks to a high memory clock rate of 28 Gbps, the GPU nevertheless achieves a very respectable memory bandwidth of 672 GB/s, which, in combination with the large L2 cache, should minimize bottlenecks.

The chip’s clock rates are designed for high efficiency. The base clock is 2,160 MHz, while up to 2,510 MHz can be achieved in boost mode. Despite these high frequencies, the targeted power consumption of 250 watts remains in the moderate range, which is made possible by an optimized power supply and energy-efficient arithmetic units.

Another key feature of the RTX 5070 is its support for modern display technologies. The GPU is equipped with three DisplayPort 2.1 and one HDMI 2.1b connection and allows resolutions of up to 8K at 165 Hz with Display Stream Compression (DSC) or 4K at 480 Hz. These features are particularly relevant for users who want to use high-resolution monitors or VR headsets with a high frame rate.

The card weighs a good 1.1 kilos (1102 grams) and its dimensions are 243 mm in length (installation length with slot), 105 mm in height (from the top edge of the slot) and 35 mm in installation depth (plus 5 mm on the backplate). An HDMI 2.1b port and three DisplayPort 2.1b interfaces are available for connecting external devices, while the card is integrated into the system via the PCIe 5.0 x16 interface.

Front

Back

Bottom

Top

Slot Panel

End

Front Bottom

ISO Back

ISO Flat Back

ISO Flat Front

ISO Front

Box Composition

Box on a Box

The GPU-Z screen shows the remaining key data of the card tested today:

I’ll save myself any further architectural descriptions at this point and refer in this regard to my very detailed launch article on the GeForce RTX 5090 as well as my assessment of raster performance and necessary paradigm shifts in graphics card development: