Things Worth Wanting / Computing & creativity

AMD Radeon RX 9070 XTThe art of making
worlds respond.

A company with character. Decades of graphics invention. And the remarkable work of making an imagined place keep up with you.

By Dr. Rob Urban

In 1979, AMD used asparagus to explain an investment strategy. The point was that some valuable things take time to grow. A collection of AMD advertisements preserved by the Computer History Museum records the analogy, and I am delighted that somebody let a vegetable into a semiconductor conversation.

Decades later, the Radeon RX 9070 XT is a very different kind of harvest. It helps build the light, surfaces, and motion of a game while a human being sits there continually changing the assignment. Look left. Run forward. Turn around because you forgot what you came for. The world has to keep up.

The company before the card

A serious business with a sense of humor.

AMD began in 1969. Its own company account describes a Silicon Valley startup with dozens of employees and ambitions in semiconductor products. The museum’s advertisement collection puts faces to the beginning: the first ad featured eight founders, including Jerry Sanders, all formerly at Fairchild Semiconductor.

Those early materials are worth a look because they contain personality as well as circuitry. David Laws, the museum curator who assembled the retrospective, had himself worked at AMD. His account records a company that hired the advertising firm Keye/Donna/Pearlstein and used wit to explain what it was building. The asparagus campaign was about investments that would need years to bear financial results.

As someone interested in how companies tell a meaningful story, I appreciate the confidence behind that choice. A good analogy respects the reader’s intelligence while giving it somewhere to land. Technical work can be difficult to explain, but difficulty does not require a personality transplant. You can discuss an ambitious business without sounding as though the quarterly report has swallowed the person writing it.

That seems especially relevant to a graphics card. The vocabulary can become a fence: compute units, memory bandwidth, ray accelerators, inference, frame generation. Each term names something real, yet a page full of them can leave the reader farther from understanding why the thing matters.

The asparagus ad offers a useful permission slip. Begin with something a person can picture. Then take the explanation seriously. In Radeon’s case, the thing to picture is an invented place that reacts when you move. The technology matters because it helps make that reaction feel coherent, immediate, and worth your attention.

AMD’s long history also gives the current card a useful sense of scale. The company did not begin with an RX 9070 XT waiting to be uncovered inside its first office. It developed capabilities, hired people, absorbed new expertise, and kept choosing what to build next. The story contains several beginnings, which is usually how an enduring technology business actually works.

A second inheritance

The graphics story joins the processor story.

One of those beginnings arrived through ATI Technologies. The Computer History Museum’s history of the companies descended from Fairchild records AMD’s 2006 acquisition of the Canadian graphics business. That is an essential part of Radeon’s identity: the graphics expertise has a history of its own, rather than appearing as a decorative attachment to a processor company.

The distinction matters. A general-purpose processor and a graphics processor have overlapping uses for arithmetic, but their organization reflects different kinds of work. Graphics offers enormous amounts of work that can happen in parallel: many pieces of geometry, many samples, many pixels, and many similar calculations. Making that work arrive efficiently at a screen is a specialty.

Bringing graphics and processor expertise into one company widened what AMD could design. The interesting part of an acquisition like this is the knowledge that becomes available to the next generation of teams. People know how to solve problems that another group has been approaching from a different direction. Their tools, habits, and accumulated lessons become part of a larger technical conversation.

A corporate family tree cannot show every contribution. The engineer who makes a scheduler more effective may never become a familiar name to the person playing a game. Neither may the people validating a display engine or helping a studio understand a performance capture. Yet the experience on the screen depends on this kind of work, repeated across product generations.

That is one reason I enjoy looking past the large badge on a graphics card. Radeon is a recognizable name attached to a changing body of work. The name carries continuity; the hardware beneath it reflects new decisions about what games demand and how a chip should answer.

AMD’s RDNA architecture history shows that evolution clearly. The first RDNA generation powered the Radeon RX 5000 series. RDNA 2 added dedicated ray-tracing acceleration and Infinity Cache. RDNA 3 brought chiplet technology to the family and introduced AI accelerators. RDNA 4 develops the computing, ray-tracing, and AI pieces further.

Those are changes in how the work is organized, not simply a sequence of model numbers. An architecture is a plan for cooperation inside the processor. It decides how resources are arranged, what they can do, and how efficiently information reaches them. A new generation becomes interesting when that arrangement fits the next set of problems better.

The people who choose the next problem

An engineer’s view from the top.

Dr. Lisa Su brings another kind of continuity to this story. Her AMD biography describes a career that moves between semiconductor engineering and business leadership. She joined AMD in January 2012 after roles at Freescale and IBM, where her responsibilities included silicon technology and research. She now serves as AMD’s chair and chief executive officer.

I find that combination compelling because a technology roadmap asks two questions at once. What can the organization learn to build? And what will be useful enough that somebody should spend years learning to build it? A technically impressive answer to only one of those questions leaves a great deal unfinished.

Graphics makes the relationship unusually visible. The chip needs to serve software that may still be under development. The software needs to reach machines with different capabilities. A new rendering technique needs enough hardware support to be practical, and the hardware needs developers to put that support to work. Progress involves coordination across schedules, teams, and products.

That is why I am cautious about giving a single executive credit for every circuit. Leadership can set direction, support difficult work, and keep an organization focused. Engineers and developers still have to solve the actual problems. The better company story makes room for both forms of contribution.

The RX 9070 XT gives us a concrete result to examine. It belongs to a generation in which traditional graphics work, dedicated ray tracing, and machine-learning-assisted rendering have all become part of the same conversation. AMD has built hardware for those tasks and software that lets game developers use it.

To me, that is a more interesting expression of ambition than a claim that everything should simply be faster. Faster at what? With which tools? Toward what experience? A product with a clear answer can be appreciated in human terms, even when the process of making it is far beyond what most of us will ever need to understand.

The person at the keyboard does not need to know how a semiconductor company is organized. They need the door to open, the camera to turn, and the light through the doorway to make sense. Getting all of that to happen together is a substantial assignment for the people on the other side of the screen.

The Radeon in this story

Meet the RX 9070 XT.

The AMD Radeon RX 9070 XT is an RDNA 4 desktop graphics processor with 64 compute units, 64 ray accelerators, 128 AI accelerators, and 16 GB of GDDR6 memory. AMD lists 53.9 billion transistors. That is a remarkable amount of organized complexity to place in the service of a person deciding whether to investigate a suspicious cave.

The number is interesting, but the division of labor is more revealing. Compute units handle broad graphics and computation tasks. Dedicated ray hardware helps with the geometric questions involved in tracing light paths. AI acceleration supports appropriate machine-learning operations. These resources give developers different ways to spend a finite amount of rendering time.

The memory has its own role. Textures, geometry, buffers, and other working data need somewhere accessible to live while the GPU constructs the scene. AMD specifies a 256-bit memory interface and up to 640 GB per second of memory bandwidth. Capacity and bandwidth describe different things: how much data can be held and how quickly data can travel through the memory interface.

This is the newer-generation Radeon chosen for this story in place of the originally considered RX 7900 XTX. It is not a simple upgrade in every specification. The RX 7900 XTX has 24 GB of memory; the RX 9070 XT has 16 GB and the newer RDNA 4 architecture. The reason to focus here is the current architectural direction, especially the ray-tracing and AI capabilities, rather than a claim that every number got larger.

There is another distinction worth keeping clear. RX 9070 XT identifies the AMD graphics model; it does not identify one universal cooler, card length, or connector arrangement sold by every board partner. AMD’s current Radeon lineup presents partner-built cards, and its own illustrative rendering is explicitly labeled as unavailable for purchase.

That lets the article stay focused on the right object. I am interested in AMD’s GPU and the technologies around it. A particular retail board adds its own physical implementation, which should be checked against its maker’s specifications when someone is building a computer.

I have not tested this card, so there are no invented frame-rate results hiding in the enthusiasm. What earns my attention is the documented design: a modern graphics processor equipped to handle several different kinds of work that converge into one visible experience. The next question is what those kinds of work actually accomplish.

Inside an ordinary-looking scene

The puddle has a surprisingly difficult job.

Imagine a game character walking past a puddle beneath a streetlight. The scene sounds simple enough. Most of us have managed something similar without a firmware update. For a rendering system, however, that small patch of water creates a collection of questions.

What is reflected in it? Which objects block the light? How does the reflected view change when the camera moves? Is the surface smooth, rippled, dirty, or partly covered? The game has to turn the answers into an image, then do the work again as the player changes position.

A great deal of real-time graphics has traditionally used rasterization: geometry is projected into the camera’s view, and the system determines which screen samples it covers. Shading then gives those surfaces their appearance. This approach remains central to modern game rendering. Ray tracing can be added to answer selected lighting and visibility questions by following rays through the scene.

Tracing every possible interaction of light would be far too much work for an ordinary game frame. A rendering system chooses which paths to evaluate, which effects to approximate, and how to reuse information. The art is in spending effort where the result matters. A convincing scene depends on the whole arrangement, not on attaching one prestigious technique to every surface.

AMD describes RDNA 4 as having third-generation ray-tracing accelerators. The significance is that certain ray-related operations receive hardware designed for them. A general calculation can be expressed in software, but dedicated support can make a useful operation less expensive to repeat. That gives the developer more room within the frame’s time budget.

The important experience is consistency. A reflection should behave plausibly when you move. A shadow should help a character seem grounded in the scene. The light should explain the shape of a space well enough that you stop consciously auditing it. When these cues agree, an invented environment becomes easier to accept.

As a writer, I recognize something familiar in that. A fictional place does not become believable because every object has been described at equal length. It becomes believable when the details support one another. A game adds the challenge that the audience can turn around and inspect the part of the room you were hoping they would politely ignore.

That is a wonderful demand to place on an invention. The graphics processor helps sustain a world under questioning. The player supplies the questions through movement; the rendering system supplies another image. The RX 9070 XT’s ray-tracing hardware is one part of how that conversation can become richer.

The constraint behind the spectacle

A beautiful frame still has a deadline.

At 60 frames per second, the average interval between frames is about 16.7 milliseconds. At 120, it is about 8.3 milliseconds. Those figures are simple arithmetic, but they give the graphics problem a useful shape. A scene can contain a cathedral, weather, moving characters, and elaborate materials. It still has to arrive on time.

There are many systems involved besides the GPU. The game’s simulation, CPU work, graphics commands, rendering, presentation, and display all contribute to the experience. A frame interval is not the same as total input-to-display latency. Still, the interval explains why the industry cares so deeply about avoiding unnecessary work.

A high average frame rate also does not tell the entire story of motion. If delivery is irregular, the player can feel the variation. Think of a sentence read aloud with unexpected pauses between otherwise familiar words. All the words may be present, but the timing changes the experience.

This is where graphics engineering becomes more interesting than a contest to perform the most arithmetic. The question is which arithmetic is needed now, which information can be reused, and which result can be reconstructed well enough to preserve the scene. The answer changes as hardware, algorithms, and game engines improve.

AMD’s GPUOpen discussion of FSR Redstone and neural rendering frames the problem around the rising cost of visual complexity. Its author, Josh Hort, describes machine learning as a way to help deliver detailed rendering within practical limits. That gives the newer AI hardware in Radeon an identifiable job rather than leaving it as an impressive label on a specification sheet.

I like this part of the story because it rewards ingenuity as much as scale. There is dignity in discovering that a difficult result can be obtained more efficiently. A good shortcut is a piece of understanding: knowledge about what information matters and how to recover what the viewer needs to see.

The ambition remains artistic as well as technical. A developer wants to keep the lighting, the materials, and the motion that give the world its identity. The rendering system needs a workable way to deliver them. What matters to the player is that the decisions preserve the feeling of the place.

The image becomes a collaboration

Rendering learns another way to work.

AMD now groups several machine-learning rendering technologies under FSR “Redstone”. The familiar FSR 4 upscaling name has become FSR Upscaling within that broader family. The change in naming makes more sense once you see that the features solve different problems.

Upscaling reconstructs a higher-resolution image from a lower-resolution rendered input, using the information available to the algorithm. It is doing more than stretching a picture until the edges reach the screen. The purpose is to recover a detailed-looking result while reducing some of the cost of producing the original frame.

That idea has an intuitive appeal if you have ever recognized a place from a small number of well-chosen clues. The rendering algorithm is not recognizing places in the human sense, but it is using structure in its inputs to build a useful output. For the viewer, the important question is whether the reconstructed image holds together in motion, where mistakes have less opportunity to hide.

Frame generation addresses a different gap: the interval between rendered frames. AMD’s developer explanation describes using successive frames, depth, motion vectors, and optical-flow information to help predict an intermediate image. The aim is smoother displayed motion. It does not mean the game’s simulation has independently produced another full state at that instant.

That distinction lets us enjoy the idea without giving it magical properties. More displayed frames and more freshly simulated, input-responsive frames are not identical claims. The underlying rendering rate and latency still matter. AMD’s developer guidance recommends pairing frame generation with its Anti-Lag 2 integration where supported to help manage responsiveness.

Ray Regeneration works on another part of the image. Ray-traced rendering may use relatively few samples to control cost, leaving a noisy estimate that needs reconstruction. Machine-learning denoising helps turn that sparse information into a cleaner result. Here the task is preserving lighting detail while making a limited set of samples more useful.

I find the three roles easier to appreciate when they stay distinct. Upscaling concerns the reconstructed image resolution. Frame generation concerns intermediate moments. Ray Regeneration concerns the quality of ray-traced information. Each approaches a different part of the same problem: delivering an engaging image with finite time and resources.

These are supported-game features, not a guarantee that every title automatically uses the entire suite. AMD publishes compatibility information because the hardware, driver, and game integration all matter. Its current page also extends ML-based upscaling to Radeon RX 7000 series cards, so the older launch-era claim that all such upscaling is exclusive to RX 9000 would be misleading today.

The RX 9070 XT is compelling as a home for this evolving work. Its AI accelerators sit beside resources for more traditional rendering, allowing different methods to contribute to the frame. The broader direction is a graphics pipeline that can calculate, trace, reuse, and infer, with developers choosing where each method helps.

The work behind the work

The people making the tools deserve a credit, too.

A feature on a chip does not walk into a game on its own. Someone has to expose it through software, explain how it behaves, integrate it into an engine, test the result, and work out what happens when the scene becomes complicated. That is a substantial creative and engineering job in its own right.

AMD’s GPUOpen developer portal is a good place to see that work in progress. Alongside rendering technologies, it provides developer tools, documentation, research, and practical guidance. The page is a reminder that the final game depends on a community with very different specialties, many of which are invisible during play.

An artist may be trying to preserve the character of a particular material. An engine programmer may be trying to reduce a bottleneck. A graphics researcher may have a new way to estimate light. A tools engineer may make it possible to understand where the milliseconds went. Their contributions meet in the same frame, even though their daily tasks look nothing alike.

This is where I think the company story becomes most tangible. AMD is doing more than placing a processor on a specification page. It is helping establish a set of possibilities that software creators can investigate. Documentation and development tools make the hardware’s capabilities easier to turn into something another person can experience.

The relationship continues after a product’s introduction. Drivers and game integrations evolve. Rendering techniques gain support or change implementation. A card is a physical object, but part of its useful life is shaped by the software around it. That makes a current account of the technology more valuable than freezing the story at the launch presentation.

For this article, it also means resisting an easy exaggeration. I cannot attribute every impressive visual effect in a game to AMD, or to any one GPU. Artists, studios, engine developers, and hardware teams contribute different parts. Radeon’s role is important enough to describe accurately without borrowing everybody else’s work.

There is a kind of generosity in good development infrastructure. The people using it can spend more of their attention on the thing they are trying to make. The best outcome may be a player who never thinks about the toolchain at all, because the world in front of them has become more interesting.

Why this Radeon belongs

Keep a record of what we asked it to make.

For my Things Worth Wanting collection, the RX 9070 XT represents a particularly interesting point in the history of graphics. Dedicated rendering hardware and machine-learning methods are being brought together to serve an experience that must continually respond to a person.

That is more specific than admiring a large transistor count. I am interested in the allocation of effort: hardware for tracing rays, resources for AI operations, memory for the working scene, and software that lets those pieces contribute. The object records a set of decisions about how to turn limited time into a believable image.

It also carries a company story with some character. AMD’s early advertisements were willing to be funny. Its graphics heritage includes ATI. Its present leadership includes an engineer with a long history in semiconductor work. Its developer community keeps working on ways to make difficult visual ideas practical. Those threads give the product depth beyond its model number.

A digital time capsule should preserve some account of our expectations, too. We asked computers to let us enter invented places, inspect them from unexpected angles, and have those places respond quickly enough that we could care about what happened there. That is an astonishingly demanding form of entertainment to regard as normal.

I would want someone reading this later to understand the pleasure as well as the engineering. A well-made game can reward curiosity. You take the longer route because a distant shape looks interesting. You stop to watch the light change. You notice the care someone put into a corner that you were never required to visit.

The Radeon RX 9070 XT helps provide the computational means for that care to reach a player. It cannot supply the artist’s idea or the player’s curiosity. What it can do is help the two meet, frame after frame, while the player keeps asking the world another question.

That is the part I want to remember: we put an extraordinary amount of invention into giving imagination a place where people could move around.

A few useful distinctions

A clearer picture.

Why feature the RX 9070 XT instead of the RX 7900 XTX?

The RX 9070 XT represents AMD’s newer RDNA 4 graphics direction, including newer ray-tracing and AI accelerators. It is not a like-for-like improvement in every specification: it has 16 GB of GDDR6 memory, while the RX 7900 XTX has 24 GB. This article focuses on the newer architecture rather than claiming a universal upgrade.

What does RDNA 4 contribute to this Radeon?

AMD identifies RDNA 4 as the architecture behind Radeon RX 9000 series graphics. The RX 9070 XT combines 64 compute units, 64 ray accelerators, and 128 AI accelerators, supporting different parts of graphics and machine-learning workloads.

Is FSR 4 still the current name?

AMD now calls that technology FSR Upscaling within its broader FSR Redstone family. The suite also includes frame generation, ray regeneration, and radiance caching. Feature support depends on the particular game, hardware, driver, and integration.

Does frame generation mean the game runs its simulation faster?

Frame generation produces intermediate displayed images between rendered frames. It does not by itself mean the game simulation is producing equally many new input-responsive states. Base rendering performance and latency remain relevant to how the game feels.

Is every RX 9070 XT retail card physically identical?

No. AMD’s graphics model appears in partner-built cards with differing coolers, dimensions, power connections, and other implementation details. Check the specifications of the exact retail board when planning a computer build.

Has Rob personally tested the RX 9070 XT?

No firsthand testing or ownership is claimed here. This is a researched editorial article about the company, architecture, and creative significance of the Radeon RX 9070 XT. Specifications are attributed to AMD, and no independent benchmark results are presented.

Sources & further reading