NVIDIA GeForce RTX 5090 vs. NVIDIA GeForce RTX 4090, Which Is Better?

NVIDIA® GeForce RTX™ 5090 vs. NVIDIA® GeForce RTX™ 4090

Even before NVIDIA CEO Jensen Huang delivered his CES 2025 keynote, there had been continued buzz surrounding the arrival of the new NVIDIA GeForce RTX 50 Series GPUs. However, nearly all of that information derived from leaks followed by a fair amount of speculation. Since the keynote, where specs, pricing, and performance (to some degree) were discussed, and the obvious fact that the new GPUS are available to order inside select BOXX systems, it’s time to see how they stack up against their predecessor, the NVIDIA GeForce RTX 40 Series Ada Lovelace GPUs, first introduced in 2022. NVIDIA calls the new GeForce GPU a “game changer” that optimizes Generative AI (more on that later) and delivers more power, memory, features, and compute.

Comparing the NVIDIA GeForce RTX 5090 vs. NVIDIA GeForce RTX 4090 in Raw Compute & Resolution Scaling

When isolating cards for pure, comparable raw compute performance without neural assistance, the generational leap is highly dependent on resolution and workload density.

• Geometric Mean Increase: Across a massive testing suite, the RTX 5090 yields a 23% overall performance uplift over the RTX 4090.
• High-Resolution Scaling: In native 4K workloads, where the GPU is fully utilized and free from system CPU bottlenecks—that performance delta climbs to a 26% uplift.
• The Compute Outlier: In maximum-stress scenarios featuring heavy path-traced rendering pipelines, the performance gap can spike significantly higher (up to 31% to 42%). However, in extremely dense compute tasks, single-digit performance swings can wildly alter percentages due to rounding metrics, so a sustained ~30% improvement in extreme ray-tracing is the safer production estimate.

To fully utilize the RTX 5090's massive parallel processing capabilities, pipelines must be optimized for dense, high-resolution outputs (such as 4K video editing, complex 3D viewports, or heavy data processing). At lower resolutions, even the fastest top-tier marketplace CPUs will bottleneck this card, leaving raw GPU performance on the table.

What are the key features of the NVIDIA GeForce RTX 50 Series?

The RTX 50 Series introduces several significant features, including DLSS 4, which enhances performance through Multi Frame Generation, allowing for the generation of up to three frames compared to one with DLSS 3. It also incorporates improvements in raw performance, with the RTX 5090 boasting a 30% increase over the RTX 4090.

Key Features of the RTX 50 Series

• DLSS 4 with Multi Frame Generation: Generates up to three AI-created frames for every rendered frame, significantly increasing frame rates compared to DLSS 3.
• Enhanced Ray Tracing: Improved RT cores provide more realistic lighting, reflections, and shadows.
• Blackwell Architecture: Delivers greater efficiency, higher performance, and advanced AI processing capabilities.
• Faster AI Performance: New Tensor Cores accelerate AI-driven workloads, gaming enhancements, and creative applications.
• Improved Memory Subsystem: Higher memory bandwidth and larger VRAM capacities support demanding games and professional workloads.
• Lower Latency Technologies: NVIDIA Reflex improvements help reduce input lag for competitive gaming.
• Next-Generation Content Creation: Enhanced support for AI-assisted rendering, video editing, and 3D design workflows.

Blackwell RTX 50-Series GPU

The NVIDIA GeForce RTX 50 Series GPU is part of the next-generation Blackwell RTX 50-Series GPU family and equipped with up to 92 billion transistors. In addition, the Blackwell streaming multiprocessor (SM) offers more processing throughput and, as NVIDIA states, “a tighter integration with the Tensor Cores in order to optimize the performance of neural shaders.” The new Tensor Cores also provide tons of AI processing power and support accelerated processing of FP4 precision models. Even better, Blackwell Tensor Cores process models faster using less graphics memory.

Key Blackwell architecture enhancements include:

• Up to 92 billion transistors
• Increased processing throughput with the new streaming multiprocessor (SM)
• Tighter integration with Tensor Cores to optimize neural shader performance
• Support for accelerated FP4 precision model processing
• Faster AI model processing with reduced graphics memory usage

Blackwell also features new RT Cores designed for ray tracing massive amounts of detailed geometry. The RT Cores have twice the ray triangle intersection rate of the previous generation, and enhanced compression designed to reduce memory footprint. Blackwell GPU architecture has also been further enhanced with PCIe Gen5 and DisplayPort 2.1b UHBR20, driving displays up to 8K 165Hz.

Additional platform improvements include:

• New RT Cores for advanced ray tracing performance
• Twice the ray triangle intersection rate of the previous generation
• Enhanced compression to reduce memory footprint
• PCIe Gen5 support
• DisplayPort 2.1b UHBR20 support for displays up to 8K 165Hz

One more item to note is that Blackwell architecture was built with what NVIDIA refers to as “enhanced hardware flip metering capabilities” to provide speed and accuracy required for a smooth, high-quality experience. That leads us to NVIDIA DLSS 4 MFG (Multi Frame Generation) and the aforementioned NVIDIA RTX Neural Shaders.

• NVIDIA DLSS 4 Multi Frame Generation accelerates frame rates by up to 8X
• NVIDIA RTX Neural Shaders which empowers creators with “next-level graphical fidelity”

DLSS 4 with Multi Frame Generation, works with a suite of DLSS technologies to multiply frame rates by up to eight times over traditional rendering and deliver superior image quality. It can generate entire frames and fully understand scene composition including shadows, reflections, and occlusion.

NVIDIA contends “RTX Neural Shaders will drive the next decade of graphics innovations,” as they save huge amounts of graphics memory by compressing textures up to 7X and offer the capability to create cinema-like textures.

RTX Neural Shader benefits include:

• Texture compression up to 7X
• Reduced graphics memory usage
• Capability to create cinema-like textures

That’s the big news on the NVIDIA GeForce RTX 50 Series, so let’s take a look at how the specifications and features compare to the previous gen NVIDIA GeForce RTX 4090.

How does the NVIDIA GeForce RTX 5090 compare to the RTX 4090 in terms of performance?

GeForce RTX 5090 GPU Cores

With 170 Blackwell Streaming Multiprocessors (SMs), GeForce RTX 5090 boasts a 33% increase in GPU cores over the 4090’s 128 SMs. Since the amount of CUDA cores, tensor cores, RT cores, texture units, etc. is directly related to the SM counts, that's an overall increase of 33%.

Clock speed

For raw compute, the 5090 (2407MHz boost clock speeds) offers a 27% improvement over the 4090 (2520MHz boost clock speeds).

GDDR7

How does GeForce RTX 5090 fuel all this processing power? Blackwell is equipped with GDDR7, or what NVIDIA refers to as “the world's fastest memory.” With speeds up to 30Gbps, Blackwell GPUs can deliver up to 1.8TB/s of memory bandwidth. More memory capacity means:

• Multi-application workflow support
• Increased bandwidth
• Improved support for AI/ML workloads
• Faster content creation

When it comes to VRAM, GeForce RTX 5090 has 33% more than the 4090, and this very substantial increase will pay big dividends for a number of tasks including:

• Next-gen raytracing & AI-powered graphics
• AI-assisted video editing/rendering
• Real-time 8K video editing
• Faster 3D rendering for cinematic production

AI and the Neural Leap: From CNN to Transformer Networks

As promised, let’s look at AI where you’ll see some major advancements. For starters, GeForce RTX 4090 has 661 TFLOPS of FP16 tensor compute and 1321 TOPS (teraops) of INT8 tensor compute. Compared to the RTX 5090, these numbers now appear downright pedestrian.

One of the most significant architectural updates in the Blackwell generation is how the GPU handles AI-driven upscaling and reconstruction. NVIDIA has completely overhauled the underlying machine learning model.

• Previous generations relied heavily on a Convolutional Neural Network (CNN) architecture to handle tensor-driven image upscaling and ray reconstruction.
• The RTX 5090 introduces the industry’s first real-time application of a Transformer model architecture.

For professional visualization, real-time virtual production, and interactive architectural walk-throughs, the Transformer model provides vastly superior temporal stability, sharper motion details, and a massive reduction in ghosting artifacts. While this model features backward compatibility with Ada Lovelace architectures, the RTX 5090's upgraded Tensor cores process this highly complex network far more efficiently.

GeForce RTX 5090 delivers up to 1676 TFLOPS of tensor FP16 compute, and double for 3352 TOPS of tensor INT8 compute for an astonishing 154% increase (2.54X) in AI computational performance.

The NVIDIA DLSS 4 relies on Blackwell’s new features to power its AI algorithms so (as mentioned earlier) multi frame generation will generate up to three additional frames from one rendered frame making everything look smoother and of higher quality.

AI transformers have been the lynchpin of the AI universe powering DALL-E, ChatGPT, and other AI content generators. NVIDIA’s DLSS Transformer upscaling relies on a newly trained network built off of AI transformers, rather than the convolutional neural network (CNN) used with earlier DLSS upscaling algorithms. Also, the new DLSS Transformer algorithm is faster than the CNN version, and will be available for all NVIDIA RTX GPUs. If you want to improve your workflow by taking advantage of all AI has to offer, upgrading to a BOXX with this new GPU is a great place to start.

NVIDIA Tensor Compute & Architecture Comparison

NVIDIA GeForce RTX 50 Series GPUs are available now inside the BOXX Creativ PC systems. To learn more about BOXX Workstation GPU choices, visit boxx.com, explore the BOXX blog, or contact a BOXX performance specialist via online chat or call 877.877.BOXX.