Things Worth Wanting / NVIDIA
NVIDIA GeForce RTX 5090: Teaching Silicon to See the Light
A puddle in a video game can contain a surprisingly large piece of computing history. Follow its reflection far enough and you reach a company founded to make better graphics, a revolution in parallel computing, and artificial intelligence finding its way back to the screen.
Start with something ordinary
That puddle has a lot to explain.
Imagine a quiet street after rain. A shop window throws warm light onto the sidewalk. A passing car changes the reflection in a puddle. You take a step sideways, and the whole arrangement shifts.
Nothing about that scene seems remarkable until someone asks a computer to make it happen.
Now the window needs a position, a surface and a source of illumination. The road needs a description of how it responds to light. The reflection has to change with the viewer. Whatever calculation made the last image convincing needs to be ready to do it again as soon as you move.
I find that transformation fascinating. An everyday visual experience becomes an enormous collection of decisions. The world has been casually doing this all morning. The computer would appreciate a little more time.
The NVIDIA GeForce RTX 5090 belongs in my digital time capsule because it represents an ambitious answer to that time problem. It also carries a much bigger story about how a machine built to draw imaginary worlds helped create tools that now reshape computing.
This is an exploration of the desktop RTX 5090, using NVIDIA’s own Founders Edition for the physical design discussion. I have not personally benchmarked or tested the card. The interest here is what its engineering reveals and how it fits into the company’s longer journey.
Following a ray
The authors of Physically Based Rendering explain the essential machinery: a camera, geometry, materials, light sources and calculations of how light travels between them. A renderer can trace viewing rays from the camera into a scene, find the surfaces they encounter and estimate the light reaching the viewer.
Additional paths account for light arriving through reflections or other interactions. A surface can receive illumination indirectly from its surroundings. That is part of why a room can look believable even where no lamp shines straight onto the wall.
The computer does not need to recreate every physical detail of the universe. It needs a useful mathematical model. For many everyday scenes, geometric descriptions of light provide a workable route to an image.
That choice of model is beautiful to me. Engineering often advances by deciding which details matter to the question. A convincing patch of sunlight does not require us to build a miniature sun inside the case.
And it leaves an interesting role for taste. The software can calculate a reflection, but someone still chooses the room, the light, the camera and the mood. Better tools give that person a richer vocabulary.
The deadline inside the picture
16.7 milliseconds is a very short workday.
At a steady 60 frames per second, a new frame arrives roughly every 16.7 milliseconds. At 120, the interval falls to about 8.3 milliseconds. Those figures are simply one second divided by the frame count.
A 3840-by-2160 image contains 8,294,400 pixels. Multiply by 60 and you have almost half a billion displayed pixel positions each second. That arithmetic is not a measure of GPU instructions or a benchmark. It is a way to appreciate how crowded the schedule becomes before we even ask what each pixel should contain.
Consider what makes a virtual room interesting: a brushed metal handle, a soft curtain, a glossy floor, a figure moving through the doorway. The player expects them to belong to the same world. A beautiful still frame is only the beginning of the challenge.
I like thinking about real-time graphics as an agreement with the viewer. The scene can be fantastical, but it has to respond coherently. Move the camera and the world should keep its promise.
Speed therefore has an artistic dimension. When the scene responds promptly, the player can explore it through movement. When an artist can inspect a change quickly, the artist can make another decision while the original idea is still fresh.
That is the sort of benefit I care about beyond a headline number. Waiting changes the rhythm of a task. Reducing the wait can change how freely a person experiments.
A company follows its own invention
Graphics opened a much larger door.
NVIDIA’s corporate history dates its founding to April 5, 1993. Jensen Huang, Chris Malachowsky and Curtis Priem began with a vision of bringing 3D graphics to gaming and multimedia.
The same timeline marks the GPU milestone in 1999, the unveiling of CUDA in 2006, the AlexNet neural-network breakthrough in 2012 and the arrival of RTX in 2018. Together, those dates sketch a business whose original specialty opened into a much broader field.
What interests me is the continuity beneath the change. Drawing a complicated image involves many calculations that can be distributed across a processor. Other people had complicated calculations too.
There is a familiar temptation in company histories to make every later achievement sound inevitable. The ending is known, so the early choices start looking like clues planted by a screenwriter. I prefer to notice how unusual the path actually is.
A business founded around gaming and multimedia eventually supplied an important foundation for modern AI. Now a consumer graphics card carries specialized hardware for neural-network work. The connection makes sense once you understand it. That does not make the journey ordinary.
CUDA made the invitation practical.
NVIDIA describes CUDA as a parallel computing platform and programming model. It gives developers a way to use supported GPUs for general computing tasks, with a surrounding system of tools and libraries.
Think of a large collection of numbers that need similar operations. If the work can be divided appropriately, a GPU can work on many parts at once. The programmer still has to organize the task and move data effectively. Some work fits that arrangement far better than other work.
My favorite analogy is a room full of workstations with a carefully prepared job at each one. The room is useful when the jobs can proceed together. If everyone must wait for one person to finish before starting the next step, adding more workstations does not solve the delay.
The software matters because it is the invitation to use the room. A powerful chip becomes much more useful when people can build dependable tools around it.
That is a company-building lesson I find compelling. A product can solve today’s problem and also become a place where other people solve problems its creators never fully anticipated. The wider contribution grows through that participation.
It is also why my interest in artificial intelligence and machine learning includes the machinery underneath the experience. The visible application is supported by years of less visible work on chips, programming tools and mathematical methods.
Meet the object
The desktop RTX 5090.
Several kinds of work share one graphics processor.
NVIDIA introduced the RTX 5090 in its January 2025 GeForce announcement, with availability beginning January 30. Its Blackwell GPU contains about 92 billion transistors. The launch specifications include 1,792 GB/s of memory bandwidth.
For the current hardware reference, NVIDIA’s RTX 5090 specification table lists 21,760 CUDA cores, fifth-generation Tensor Cores, fourth-generation ray-tracing cores and 32 GB of GDDR7 memory on a 512-bit interface. The company’s current GeForce catalog continues to list the desktop RTX 5090.
Those labels describe different resources. They are not interchangeable counters that should be added together to produce one grand score.
NVIDIA’s Blackwell architecture white paper describes a design combining programmable shader processing, specialized ray-tracing work and Tensor Core operations. Blackwell also supports FP4 computation for compatible AI workloads, using a compact numerical representation.
The useful idea is specialization. Some calculations benefit from flexible programmable machinery. Others happen often enough in a recognizable form that dedicated hardware makes sense. The overall experience depends on the parts working together.
I appreciate that kind of design because the visible result can be wonderfully simple. A player sees a reflection. Behind it, an entire organization of computational labor has been assembled.
The impressive part is the coordination. A collection of fast components still needs the right information at the right moment. The experience on screen is the result of the system.
A number with practical meaning
Memory is room to work.
The 32 GB memory figure is easier to understand if we separate capacity from bandwidth. Capacity describes how much data can be held. Bandwidth describes a rate at which data can move. A large desk and a fast delivery system solve different problems.
In graphics, that working space may hold textures, geometry and intermediate results. In an AI task, it may hold model data and the additional information needed while the model runs. Exactly what fits depends on the software and workload.
Imagine preparing a complicated illustration with reference material spread across a table. A bigger table gives you more room to keep things close. It does not automatically make every decision faster, but it can change whether the working arrangement is comfortable at all.
That is why I resist turning memory capacity into a universal performance promise. It is a meaningful resource with a particular job. Its value becomes clearer when the task is named.
For someone exploring local AI, the distinction between learning and using a learned model matters too. NVIDIA’s explanation of training and inference describes training as the process that develops a model and inference as applying the trained model to new input.
Running a suitable model on a personal computer is a different undertaking from training a vast foundation model. The 5090’s memory and supported numerical formats create possibilities, but the model, precision, context and software determine the actual requirements.
I like the personal scale of that experimentation. There is something satisfying about an idea becoming a thing someone can investigate at a desk. A graphics card can be part of a private workshop for images, code, video and mathematical curiosity.
That workshop still rewards judgment. A tool becomes useful through a question worth asking and a person willing to examine the result. The silicon does not supply that part of the project.
The story bends back toward the screen
AI returns to the world that helped it grow.
Here is the loop that makes the RTX 5090 especially interesting to me: graphics helped create demand for parallel processors; those processors became useful for neural networks; neural networks are now part of producing graphics.
It is a remarkably compact company story to find inside one object.
NVIDIA’s DLSS overview describes a family of related techniques. Super Resolution reconstructs a higher-resolution image from lower-resolution inputs. Ray Reconstruction uses AI to improve the reconstruction of ray-traced imagery. Frame-generation techniques produce additional displayed frames.
The software has continued developing since the card’s launch. NVIDIA’s current overview describes DLSS 4.5 features and DLSS 5’s developer-tuned, 3D-guided neural rendering for lighting and materials. Which features can be used depends on the game, hardware and software support. Buying a card does not activate every technique in every title.
For the time capsule, that distinction matters. The physical object has a release history. Its software capabilities have a continuing history of their own.
A smoother picture and a fresher response are different questions.
Suppose an application renders one sequence of frames and a supported technique inserts additional generated images. The display can receive a denser sequence. That does not mean the underlying game simulation has independently advanced for every inserted image.
I would therefore want more context than a single frame-rate headline. How was the image produced? What settings were enabled? How promptly did a new input affect what appeared? Were moving edges and fine details stable?
Those questions make the achievement easier to understand. They also give the reader a way to interpret demonstrations without needing to become a graphics programmer.
My interest is in the creative possibility of combining methods. An engineer can allocate computation differently when a reconstruction technique handles some of the visible result. An artist gains another tool, along with another set of choices to evaluate.
There is a larger cultural question here too. When people say an image was “rendered,” they may imagine one straightforward process. Modern graphics can involve multiple stages, different numerical representations and learned models. The word still fits, but the workshop behind it has become more elaborate.
I want this collection to preserve some of that workshop. Otherwise we keep the impressive object and lose the reason it was a turning point.
The physical world collects its share
All that imagination needs airflow.
The Founders Edition makes the physical side of the story visible. NVIDIA’s launch account describes a two-slot card, a three-piece circuit-board arrangement, a liquid-metal thermal interface and a double flow-through cooling design.
The current reference specification lists a 304 mm length, 137 mm width and 575 W Total Graphics Power. Partner cards can have different dimensions and cooling designs. The laptop GPU bearing the 5090 name is a separate product.
A power specification is an engineering condition, not a statement that the card consumes that amount continuously. Actual demand varies with operation. System requirements also depend on the rest of the computer.
I find cooling design an excellent antidote to the idea that computing is somehow weightless. The image may depict a winter forest. The machine still has heat to move.
That gives the enclosure, fans, clearances and electrical connections a real role in the product. The elegant object cannot be understood completely apart from the space where it operates.
For an actual build, I would work from the exact card and power-supply documentation, with room for airflow and the specified cable clearances. “Two slots” describes the card’s form; it is not a complete installation plan. A capable builder earns their keep by checking the whole arrangement.
The same principle applies to performance expectations. The processor, application, memory, display and settings all participate. Putting an ambitious component into a computer is the beginning of a system decision.
There is something reassuring about that. Even a product surrounded by extraordinary numbers still answers ordinary engineering questions: Does it fit? Can it get what it needs? Can it shed heat? Can the person using it accomplish the intended work?
The collection’s reason for keeping it
A machine for making another attempt.
In my work, I care about the gap between an idea and something another person can experience. That gap can be frustratingly large. A person may know the mood they want, the space they imagine or the story they want to tell, while the tools still demand many intermediate steps.
The appeal of capable graphics hardware is partly the possibility of shortening that distance. Try a different light. Move the camera. Examine another material. See what the change does while you still remember why you made it.
That is a useful way to understand innovation: look at the new actions a tool makes practical. A specification becomes meaningful when it changes the work available to a person.
The RTX 5090 is worth including because several stories meet in it. There is the company that began with 3D graphics. There is the broader usefulness of parallel processing. There is the return of AI to image creation. And there is a physical object that has to carry those ambitions inside a computer.
I do not need to call it the right purchase for everyone to find it culturally interesting. A time capsule can include an object because it reveals what people were trying to achieve at the edge of a field.
Here, people were trying to make interactive worlds more convincing and creative work more immediate. They were also renegotiating how much of an image should be calculated directly, reconstructed from evidence or produced through a learned model.
Those are substantial questions to attach to something sold as a graphics card.
A few useful distinctions
Questions about the RTX 5090.
Which RTX 5090 does this article cover?
The desktop NVIDIA GeForce RTX 5090. The physical design discussion uses NVIDIA’s Founders Edition. Partner cards can differ, and the GeForce RTX 5090 Laptop GPU is a separate product.
What are the core hardware specifications?
NVIDIA lists the Blackwell architecture, 21,760 CUDA cores, fifth-generation Tensor Cores, fourth-generation ray-tracing cores and 32 GB of GDDR7 memory for the desktop RTX 5090.
Why is CUDA important to the company’s story?
CUDA gave developers a practical platform for using supported NVIDIA GPUs for parallel computing beyond graphics. That wider use helps explain the connection between gaming hardware, scientific computing and modern AI.
Does generated frame rate equal game simulation rate?
No. Additional generated images can increase the number of displayed frames without making the game simulation independently advance for each one. Rendering rate, responsiveness and visual stability are useful separate measures.
Does every game support every DLSS feature?
No. Availability depends on the game, GPU and software support. DLSS is an evolving family of technologies, so current compatibility information matters more than assuming every feature comes with the card.
Can 32 GB of graphics memory run any AI model?
No. Requirements depend on model size, numerical precision, context and the software’s working memory. Running a suitable trained model is also different from training a large foundation model.
Has Rob personally benchmarked the RTX 5090?
No. This article is an editorial exploration of NVIDIA’s history, graphics science and the desktop RTX 5090’s documented design. It does not report firsthand testing or measured performance.