AMD's next-generation GPU architecture, RDNA 5, may be announced in late 2027. The hardware leaker Kepler_L2 said on AnandTech Forums that the earliest possible window is Computex 2027, but that a date later in the year is more likely. AMD has not announced any delay. Still, if the prediction holds, buyers will be using RDNA 4 for longer. Meanwhile, compiler support for the next generation and AI rendering features on current Radeon cards keep advancing, so whether RDNA 5 is worth waiting for is hard to judge from the launch date and chip performance alone.
Read the release-timing prediction before talking about a delay
The discussion began with a user asking whether AMD might reveal RDNA 5 cards within the next three to five months. Kepler_L2 said no. When another user asked, "So June 2027?", Kepler_L2 replied that Computex 2027 would be the earliest opportunity. The reply itself favors a later point in the year.
The question was about when the cards would be "revealed," not when they would go on sale in stores, and the answer did not name a release date. The late-2027 outlook cannot simply be turned into a schedule in which every model launches in the second half. Nor does the reply say whether the top models will come first or whether only a limited set of products will ship early.
Calling it a delay would also require an earlier plan to compare against. What was shared here is a leaker's outlook, not information that AMD changed a publicly announced launch date. Even if the timing feels later than earlier predictions, nothing here explains whether a development problem occurred or whether AMD changed its product-launch decisions.
The preparation for the next generation can, however, be traced through public information. Apart from schedule rumors, looking at which designs AMD's software is starting to support makes what lies ahead a little more concrete.
What public code shows about RDNA 5's design
The official AMDGPU documentation for the LLVM compiler infrastructure already lists "GFX13 (RDNA 5)." The target identifier is "gfx1310," classified as a dGPU, meaning a discrete GPU. The product-name field, however, is still listed as TBA, or unannounced. The development name and target are confirmed, but the Radeon product name and launch timing are not.
LLVM translates programs into the instructions a GPU executes. When a new target appears in public code, it means preparation is under way to generate instructions for that GPU and run software on it. The code alone cannot tell us whether actual silicon is finished or what its production status is. Even so, RDNA 5 is no longer just a name used in rumors.
There are also clues about how computation will be used. LLVM's instruction-handling code contains a GFX13-specific branch for VOPD, which pairs two compatible instructions. In addition, the instruction definitions explicitly define a GFX13 VOPD3 format, separate from the conventional GFX12 format. The same format is also defined for a different target, GFX1250, so it is not a feature exclusive to RDNA 5.
According to the code that describes VOPD3's constraints, this format relaxes the condition on whether the register numbers holding results are even or odd, and it can also handle an independent third input. A register is a place where values used in computation are held temporarily. By combining the instruction-format definition with the constraints applied to that format, we can see what kinds of processing the GFX13 code is meant to handle.
This is not about adding more arithmetic units. If the combinations of instructions that can be processed simultaneously are narrow, the units cannot be fully used even when they have spare capacity. Widening the freedom of combinations means more situations in which the same computing resources can be put to use. How much this helps in practice, though, depends on the instruction streams of games and the combinations the compiler produces.
It is also a mistake to read the "3" in VOPD3 as three instructions issued at once. The code handles two components, X and Y, and support for a third input is separate from the number of instructions. Development changes like these cannot be tied directly to performance predictions such as "games will run twice as fast." When assessing next-generation GPUs, we need to ask not only about the number of compute units and theoretical throughput, but also in which kinds of workloads the arithmetic units can be used effectively.
RDNA 4 keeps changing while you wait
AMD also expanded its current generation at Computex 2026. In its official announcement on May 31, the company set the global launch of the RDNA 4-based Radeon RX 9070 GRE for June 2. Even as the next generation is prepared, efforts to broaden the current products continue.
Improvements to rendering features run on yet another timeline. In AMD's FSR SDK 2.3, machine-learning upscaling targets the RX 7000 and 9000 series, while machine-learning frame generation, Ray Regeneration, and Radiance Caching target the RX 9000 series and later.
| FSR SDK 2.3 feature | Role | GPUs AMD lists as supported |
|---|---|---|
| Upscaling 4.1.1 | Reconstructs a high-resolution image from a low-resolution one | Discrete GPUs in the RX 7000/9000 series |
| Frame Generation 4.0.1 | Generates intermediate frames from the images before and after | RX 9000 series and later |
| Ray Regeneration 1.2.0 | Reduces noise left in ray-tracing results | RX 9000 series and later |
| Radiance Caching 0.9.0 | Uses learned lighting information to reduce ray-tracing computation | RX 9000 series and later. Technology preview |
The table lines up each machine-learning feature using the feature descriptions and "Requirements" in GPUOpen's FSR SDK 2.3, checked on October 13, 2026. It is not a comparison of speed or image quality. Using these features requires integration on the game side, and there are also software requirements such as Shader Model 6.6. We have kept them separate from the conventional upscaling and frame generation that work on older generations.
An RX 7000 owner does not necessarily have to buy a new card to get machine-learning image reconstruction. But being able to use that is a different matter from being able to use machine-learning frame generation or ray-tracing denoising. Judging support by the single name FSR makes it easy to overlook this difference.
There are also conditions on how these features reach games. AMD explains that DirectX 12 games that have integrated FSR 3.1 or later can be updated to machine-learning upscaling for the RX 7000/9000 series through a future Adrenalin driver. Machine-learning frame generation targets DirectX 12 games that have integrated FSR 3.1.4, together with the RX 9000 series. Even with a supported GPU, not every game will immediately switch to the new features.
In other words, image quality and features are not fixed while you wait for RDNA 5. There is room to check which improvements reach current GPUs, then work out which performance and features can only be had by buying a new chip.
Measuring next-gen GPUs by light and AI processing time
AMD's developer documentation for Radiance Caching shows that light-rendering computation and machine learning already work together within a single rendering process. In each frame, a small number of light paths are actually traced, and the results are used to gradually train a neural network. At the same time, the mechanism uses the accumulated lighting information to skip unnecessary ray tracing.
The more finely the process of light hitting a surface and bouncing off another is traced, the heavier the computation becomes. Radiance Caching replaces part of that with learned predictions, reducing the average depth traced. If the lighting in a scene changes, learning continues. It differs from Ray Regeneration, which removes noise from the finished image afterward, in where it reduces the load.
Still, Radiance Caching remains a technology preview and should not be treated as a finished feature usable across a wide range of games. Nor do these materials show any dedicated circuitry in RDNA 5 or a performance gap over current GPUs. What they show is a direction in which the ability to trace rays and the ability to reduce how much is traced through learning and inference complement each other within rendering.
The further things move in this direction, the harder it becomes to measure a next-generation GPU's value by compute-unit count alone. Even if ray tracing is fast, if image reconstruction or learning takes time, that burden remains in the time it takes to finish drawing the screen. Conversely, if some computation is reduced but image quality breaks down, it is hard to call that a performance gain under the same conditions. Comparison requires checking frame times with each feature enabled, in the same scene and at the same quality settings.
What the official RDNA 5 announcement will test, alongside the release date, is how fast and how stably these processes run in real games. If the new generation can handle work current cards cannot, with visible image-quality gains and short latency, the reason to wait for it becomes much more concrete.