NVIDIA GeForce RTX 5090 Founder's Edition Review - Top of the Line Blackwell

For NVIDIA there is no turning back, and it has decided that the new architecture Blackwell For graphics cards, it employs more and better techniques and tools based on artificial intelligence, neural network processing, and generative AI engines. The green company seeks to further employ its expertise in business AI, without neglecting the classic improvements in base specifications, energy efficiency, compatibility with new codecs and video outputs, a new type of GDDR7 memory, among others.

Anyway, the star in Blackwell is the Deep Learning Super Sampling technology, popularly known as DLSS, which in its fourth generation promises up to twice as many frames per second as DLSS 3 or 3.5. There will also be a deeper look at “neural shaders,” which will work with AI models trained by developers to generate “approximate” images even faster than previous-generation ray tracing. We also have an update to the DLSS Ray Reconstruction technique, which reduces the number of rays needed to generate ray-traced lighting, and another group of techniques that we will see later.

Let's start with the basics, the specifications of the graphics cards that we will show below.

La GeForce RTX 5090 will have the GPU GB202. It includes 12 graphics processing clusters (GPCs), 96 texture processing clusters, 192 streaming multiprocessors (SMs), and a 512-bit memory interface with 16 32-bit memory controllers. Each of the 12 GPCs mentioned above offers a rasterization engine, 8 texture processing clusters, 16 streaming multiprocessors, and 16 rasterization operation partitions.

In total, the GB202 GPU has 128MB of L2 cache, although 5090MB is offered for the RTX 96. This hardware will be used to manage the architecture based on “neural rendering”. The main attraction will be the use of DLSS 4, with multiple frame generation and lower latency through the use of improved RTX techniques (SLR 2) and image generation or transformation using AI.

Finally, each streaming multiprocessor in Blackwell (SM) includes 128 CUDA cores, a fourth generation RT core, Four fifth-generation tensor cores, 4 texture units, a 256KB register file, and 128KB of shared memory or L1 cache. The FP32 and INT32 cores have also been unified, allowing them to perform one of the two operations when needed. Additionally, the number of texture units has been increased from 512 on the RTX 4090 to 680 for the RTX 5090. Bilinear texel rates have increased from 1290.2 Gigabytes per second on the 4090 to 1636.76 Gigabytes per second on the 5090.

One of the most striking points is, without a doubt, the debut of GDDR7 graphics memory in NVIDIA gaming GPUs. After two generations of cards with GDDR6X, NVIDIA will be launching this generation of memory with no less than 32GB for the RTX 5090, 16GB for the RTX 5080 and RTX 5070 Ti, and 12GB for the RTX 5070. 

Key improvements to GDDR7 include PAM3 signaling technology that promises improved signal-to-noise ratio and doubles the density of independent channels, resulting in greater memory bandwidth and improved energy efficiency.

Fifth-generation Tensor Cores now support the FP4 data format, which requires less than half the memory thanks to a compression method with virtually zero loss in quality compared to FP16. For example, an RTX 5090 with FP4 can generate images in less than five seconds using a FLUX.dev model, which would take an RTX 4090 with FP16 around 15 seconds.

Blackwell’s fourth-generation RT cores double the throughput of ray tracing test intersections, which is a high-frequency operation when generating RT images. Therefore, ray-traced frames are expected to be generated at a higher speed. Also included in this generation is a micromapping opacity engine, which helps reduce the amount of computational calculations in alpha shaders. Other dedicated ray tracing technologies include the use of mega geometry, a triangle group intersection engine, and linear scanning spheres for rendering fine geometry objects such as hair.

“Mega geometry” aims to increase geometric detail in applications that employ ray tracing. In particular, for graphics engines such as Unreal Engine 5 that employ more advanced level-of-detail systems, it allows the use of ray tracing with greater fidelity, improving the quality of shadows, reflections and indirect lighting.

Mega geometry is expected to be available on all DirectX12, Vulkan, and OptiX 9.0 APIs, and will be available on all RTX graphics cards from the Turing architecture or the RTX 20 generation.

On the other hand, we have Shader Execution Reordering (SER) 2.0. This technology allows the GPU's parallel threads to be reorganized for greater processing efficiency. This helps lighten the load on ray tracing processes such as divergent memory access and path tracing, and send information to the tensor or shader cores. There are already games and APIs that take advantage of this technique in ray tracing, and the new version promises better results.

The AI ​​Management Processor (AMP) is a context-programmable scheduler on the GPU that improves process scheduling in Windows, reducing the contextual load on the GPU. Using a RISC-V processor, AMP works with the Windows architecture to reduce latency and improve memory management, reducing CPU load for task scheduling and helping to reduce bottlenecks, while improving frame rates and multitasking in Windows. 

And on the graphics rendering front, we finally come to DLSS 4, the crown jewel at Blackwell. NVIDIA promises multi-frame rendering with improved performance and lower memory usage than previous versions of the technology, as well as improvements to previous DLSS techniques such as frame generation, ray tracing, super resolution, and deep learning anti-aliasing.

Combining hardware, architecture, and software improvements, they promise 40% faster frame rates than DLSS 3, using 30% less video memory, and a model that only needs to run once per frame. Optical flow frame generation is now AI-driven rather than dedicated hardware, also reducing the cost of frame generation and integration. A Flip Metering system shifts frame rate logic to the display engine, allowing the GPU to improve the accuracy of display sample times.

DLSS Super Resolution (SR) boosts performance by using AI to produce higher resolution frames from a lower resolution input. DLSS samples multiple low-resolution images and uses motion data and feedback from previous frames to construct high-quality images. The final product of the transformer model is more stable, with less ghosting, more motion image detail, and improved anti-aliasing compared to previous versions of DLSS. 

Ray Reconstruction (RR) improves image quality by using AI to generate additional pixels in ray-traced intensive scenes. DLSS replaces manual denoisers with an AI network trained on NVIDIA supercomputers that generates higher quality pixels between sampled rays. In ray-traced intensive content, the transformer model for RR further improves quality, especially in scenes with challenging lighting. In fact, all common artifacts of typical denoisers are significantly reduced. 

Deep Learning Anti-Aliasing (DLAA) provides improved image quality using an AI-based anti-aliasing technique. DLAA uses the same super-resolution technology developed for DLSS, constructing a more realistic, high-quality image at native resolution. The result provides greater temporal stability, motion details, and smoother edges in a scene.

Neural shaders are a technology NVIDIA is looking to introduce at Blackwell, unifying traditional shaders with the use of AI in parts of the rendering process, partially at first, and believed to be fully so in the future. Tensor cores are now accessible to graphics shaders combined with scheduling optimizations in SER 2.0 (Shader Execution Reordering) so that AI graphics with neural filtering capabilities and AI models, including generative AI, can run simultaneously in next-generation games.

Neural shaders allow us to train neural networks to learn efficient approximations of complex algorithms that calculate how light interacts with surfaces, decompress super-compressed video, predict indirect illumination from limited real data, and approximate subsurface light scattering. The potential applications of neural shaders are still largely unexplored, meaning that new applications may be found.

Among the other integrated techniques are the “RTX Neural Materials”. AI is used to replace the original mathematical model of a material or texture with an approximation, promising “cinema-quality” frames at gaming-grade speeds while using less video memory and fewer graphics card resources.

Another technique is RTX Neural Texture Compression, which leverages neural networks accessed through neural shaders to compress and decompress material textures more efficiently than traditional methods. Then there is the so-called “Neural Radiance Cache” (NRC). This uses a neural shader to cache and approximate brightness information. Taking advantage of the By learning a neural network, complex lighting information can be stored and used to create high-quality global illumination and dynamic lighting effects in real time. 

From there we have RTX Skin, a technique that NVIDIA developed based on subsurface scattering for rendering cinema images. This allows light passing through certain skin surfaces that are not entirely solid to be rendered subtly or intensely, depending on the requirements of the game, using ray tracing.

And RTX Neural Faces creates a rasterized face that is overlaid with a rough 3D layer using a generative AI model, which can infer natural-looking face results.

For video and encoding features, there’s support for chroma-sampled 4:2:2 video, which has lower data requirements than the 4:4:4 standard, making it suitable for generating HDR content. The ninth-generation NVENC encoder also supports higher-quality AV1 and HEVC video. The RTX 5090 graphics card supports up to three encoders and two decoders. A sixth-generation NVDEC hardware decoder is also available.

Finally, Blackwell has support for DisplayPort 2.1b with up to 80 Gbps of bandwidth. This will allow up to 165hz refresh rate in 8K resolution, and no less than 420hz in 4K resolution.