A new leak dropped on September 4, 2026 claims AMD is chasing GPU clock speeds as high as 3.5 GHz on its next-generation RDNA 5 architecture, a jump that would put the chipmaker’s graphics silicon in territory no consumer GPU has touched before. The figures, first reported by TechPowerUp and echoed by outlets including Club386, Wccftech, and TwistedVoxel, trace back to a single forum post from leaker Kepler_L2 on the AnandTech forums, cited originally by shattered.io.
None of this is confirmed by AMD. But the numbers are specific enough, and consistent enough across independent write-ups, that they’ve become the dominant talking point in PC hardware circles this week. Here’s what the leak actually says, how it stacks up against AMD’s last four GPU generations, and why a clock-speed rumor about a product that likely won’t ship until 2027 is already moving the conversation around Nvidia’s next moves too.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What the New AMD RDNA 5 Leak Actually Claims
The core claim is straightforward: AMD’s internal targets for RDNA 5 GPU clocks sit in a 3.1 GHz to 3.4 GHz range, with rumor tables circulating online pushing a ceiling of 3.5 GHz for what would be an entry-level SKU. TechPowerUp’s headline puts the number plainly, stating RDNA 5 “reportedly targets clock speeds of up to 3.5 GHz.” Wccftech frames the same figures differently, noting the 3.1-3.4 GHz band would sit nearly 14% above the Radeon RX 9070 XT’s out-of-the-box boost frequency of 2.97 GHz. TweakTown’s own September 2026 reporting lines up with those numbers, independently citing the same 3.1-3.4 GHz leaked clock range and a 3.5 GHz boost ceiling for the entry-level part.
That 14% gap matters because it’s the first concrete, named-outlet estimate of how much faster RDNA 5 clocks could run relative to AMD’s current flagship gaming GPU. It’s also a reminder that clock speed alone doesn’t equal performance; IPC (instructions per clock), architecture changes, and memory bandwidth all factor into real-world frame rates. Every outlet covering the story, including OC3D, has been careful to flag that AMD has not confirmed any RDNA 5 clock specifications, CU counts, memory configuration, or launch date.
Who Is Kepler_L2, and Why This Leak Is Getting Traction
Kepler_L2 is a known hardware leaker on the AnandTech forums whose posts have circulated through the GPU rumor mill for several product cycles. According to shattered.io’s September 4 writeup, the leaker’s latest post describes internal AMD material targeting the 3.1-3.4 GHz clock band, with derivative spec tables pushing individual SKU ceilings up to 3.5 GHz. The distinction matters: the 3.1-3.4 GHz figure is described as the broad target range for the RDNA 5 lineup, while 3.5 GHz is a rumored maximum for a specific, lower-tier part rather than the flagship.
Leaks like this gain credibility not from any single source but from convergence. When TechPowerUp, Club386, VideoCardz, Wccftech, OC3D, and TwistedVoxel all independently write up the same 3.1-3.4 GHz range within the same news cycle, it signals the underlying forum post is being taken seriously by outlets that have tracked AMD’s roadmap accuracy over past generations. That doesn’t make the numbers true. It does make them the current consensus rumor.
Breaking Down the Rumored SKU Lineup
Alongside the clock figures, rumor tables tied to the same leak reference a set of RDNA 5 SKU names: RX 10090 XT, RX 10070 XT, RX 10060 XT, and RX 10050 XT, with speculative maximum clocks of 3.0 GHz, 3.2 GHz, 3.4 GHz, and 3.5 GHz respectively. Those tiers track an earlier round of rumors: French outlet Generation-NT reported back in July 2025 that AMD’s flagship RDNA 5 die could land around 96 compute units on a 384-bit memory bus, with a midrange configuration closer to 64 CUs and an entry-level part around 32 CUs, roughly the same tiering structure that later spec tables would go on to echo. TechSpot added further granularity to those lower-tier parts in September 2026, describing a low-tier RDNA 5 die with 24 compute units, 3,072 cores, a 128-256-bit bus, and 8-16 GB of VRAM, alongside an entry-level die with just 12 compute units, 1,536 cores, a 64-bit bus, and the same 8-16 GB VRAM range. Outlets covering the tables stress that none of these names or per-SKU numbers are confirmed AMD product names or specifications. They’re plausible extrapolations built on top of a single leak, not a roadmap AMD has published.
| Rumored SKU | Reported Max Clock | Rumored CU Tier | Rumored Memory Bus |
|---|---|---|---|
| RX 10090 XT (flagship) | ~3.0 GHz | 96 CUs (Navi 51 / AT0 die) | 384-bit (some reports cite 512-bit) |
| RX 10070 XT | ~3.2 GHz | 64 CUs | 256-bit |
| RX 10060 XT | ~3.4 GHz | 32-48 CU range (unconfirmed) | 128-256-bit (unconfirmed) |
| RX 10050 XT (entry) | ~3.5 GHz | 32 CUs | 128-bit |
Note the inverse relationship: the smallest, cheapest chips in the rumored lineup carry the highest clock ceilings, while the flagship runs comparatively conservative. That pattern is normal in GPU design. Smaller dies with fewer compute units typically clock higher because they generate less heat per square millimeter and don’t need to balance power budgets across as many transistors. It also explains why the “3.5 GHz” headline number, while technically accurate to what’s being reported, doesn’t describe the part gamers building a high-end rig would actually buy.
RDNA 5 vs RDNA 4: What’s Actually Changing
Clock speed is only part of the RDNA 5 story. A TweakTown report from August 2025 cited leaks suggesting RDNA 5 could deliver a 5-10% IPC (instructions-per-clock) improvement over RDNA 4 at matched clock speeds, meaning any gains from higher frequency would stack on top of, not substitute for, architectural efficiency gains. Separately, PCGamesN cited leaker claims from June 2025 suggesting a roughly 20% raster performance gain per compute unit versus the prior generation, along with as much as double the ray tracing throughput, a figure TechSpot corroborated that same month, citing sources pointing to roughly 20% higher raster performance and 2x ray tracing throughput for UDNA/RDNA 5 versus RDNA 4. Both claims originate from unnamed leaker sources and should be read as predictions, not confirmed AMD data.
A Bigger Architectural Break Than RDNA 4 Was
Coverage from VideoCardz and Wccftech, both updated in November 2025, relays industry chatter describing RDNA 5 as an entirely new architecture design, in contrast to RDNA 4, which is characterized in that same reporting as closer to an incremental fix on top of RDNA 3. If accurate, that would mark RDNA 5 as AMD’s biggest graphics architecture shift since the jump from GCN to RDNA in 2019. Wccftech also notes it remains unclear whether AMD will keep the RDNA 5 name at all, given the parallel UDNA branding rumors that have circulated since late 2024.
Compute Units, Memory, and the AT0/AT2 Die Split
Separate from the clock-speed leak, spec tables reported by PCGuide in July 2025 pointed to gaming RDNA 5 chips scaling as high as 154 compute units, with non-gaming variants reportedly reaching 184 CUs. The same report suggested top-end gaming cards could carry as much as 36 GB of GDDR7 memory, a significant jump from the 8 GB to 16 GB range common on today’s midrange cards, with cheaper RDNA 5 SKUs starting around 12 GB.
NotebookCheck’s July 2025 coverage, drawing on Moore’s Law Is Dead sources, described the flagship die under the codename Navi 51 carrying 96 CUs on a 384-bit memory bus, with performance pitched somewhere between an RTX 4090 and an RTX 5080. OC3D’s write-up the same day used similar language, describing the flagship as a “50% jump” over RDNA 4 in raw compute resources. A later HardwareZone report from August 2025, however, put the same AT0 flagship die at a wider 512-bit memory interface rather than 384-bit, an inconsistency that underscores how much these numbers still shift between sources. PCGamesHardware reported in January 2026 that there could be 5 Navi 5X GPUs, including 96-CU and 64-CU designs, with the 96-CU Navi 5X carrying 6,144 shaders and one RDNA 5 design using a 384-bit memory interface. TechSpot added a fresh data point in June 2026, reporting that the RDNA 5 flagship, referenced under the codename Navi 5X, may pack as many as 12,288 cores running at up to 3.4 GHz, a core count that had already surfaced in industry speculation shared on LinkedIn back in November 2025, which similarly pointed to around 12,288 cores alongside VRAM estimates in the 24 GB to 32 GB range. Memory bandwidth projections have shifted too: Tom’s Hardware reported in September 2025 that AMD’s next-generation graphics silicon, then still discussed under the UDNA label, could pair with GDDR7 memory running as fast as 32 Gbps, well above what today’s RDNA 4 cards ship with.
TSMC N3P: The Manufacturing Node Doing the Heavy Lifting
Clock speed increases of this size rarely come from architecture alone; they typically require a smaller, more efficient manufacturing process. RDNA 4, which powers today’s Radeon RX 9070 XT, was built on TSMC’s N4P node. Even the node itself has been a moving target in the rumor mill: Tom’s Hardware reported in September 2025 that AMD’s next-generation architecture, at the time still framed as UDNA, might instead land on TSMC’s N3E process, a related but distinct 3nm variant from the N3P node that more recent leaks now point to. NoobFeed reported in March 2026 that RDNA 5 may use TSMC 3nm manufacturing, and a June 2026 TweakTown report, based on leaked roadmap information, states RDNA 5 is expected to move to TSMC’s N3P node, introducing what the leak describes as “Neural Arrays,” “Radiance Cores,” and “Universal Compression” features, reportedly co-engineered in part with Sony and Microsoft for their next console generations. That console tie-in got more concrete in August 2026, when Tech-Insider reported that Microsoft’s next-generation Xbox “Magnus” SoC is expected to use a 70-CU RDNA 5 die with 68 CUs enabled and at least 24 MB of L2 cache.
That same TweakTown report claims the node transition alone could deliver up to 18% faster clock speeds, 36% lower power draw, and a 24% smaller die area versus equivalent RDNA 4 configurations. A follow-up piece from Tech-Critter in July 2026 frames the N4P-to-N3P jump as historically associated with roughly 15-20% compute efficiency gains at equivalent clocks, and names a rumored flagship, the RX 10900 XT, as the part expected to tape out first on the new node. Both reports are explicit that this remains rumor-based roadmap material, not an AMD announcement.
Historical Context: How AMD’s GPU Clock Speeds Have Climbed
A 3.1-3.5 GHz target sounds extreme until it’s placed against AMD’s own generational trend. Every RDNA generation since 2019 has pushed boost clocks higher, partly through node shrinks and partly through power and thermal design improvements. RDNA 4’s Radeon RX 9070 XT already ships with a factory boost clock around 2.97 GHz, itself a record for a GPU built on a monolithic, non-chiplet design in AMD’s modern lineup.
| Generation | Flagship Card | Official Boost Clock | Manufacturing Node |
|---|---|---|---|
| RDNA 1 (2019) | Radeon RX 5700 XT | ~1.905 GHz | TSMC N7 |
| RDNA 2 (2020) | Radeon RX 6900 XT | ~2.25 GHz | TSMC N7 |
| RDNA 3 (2022) | Radeon RX 7900 XTX | ~2.5 GHz | TSMC N5/N6 |
| RDNA 4 (2025) | Radeon RX 9070 XT | ~2.97 GHz | TSMC N4P |
| RDNA 5 (rumored) | Unconfirmed (RX 10090 XT?) | 3.1-3.5 GHz target (leaked, unconfirmed) | TSMC N3P (leaked, unconfirmed) |
Viewed this way, the leaked RDNA 5 range isn’t an outlier jump so much as a continuation of a trend line AMD has held for three straight generations. What would be new is closing in on 3.5 GHz territory that, on the CPU side, took AMD’s Ryzen line more than a decade to approach. GPUs and CPUs scale differently, but the comparison illustrates just how aggressive the leaked figures are relative to where consumer graphics silicon has historically operated.
RDNA 5 or UDNA? The Branding Confusion Explained
A November 2024 TechPowerUp report, based on early leaks, argued AMD might skip the RDNA 5 name entirely and instead launch its next architecture under the UDNA banner, unifying its gaming and data center GPU lines under one architecture for the first time. That report implied RDNA 5 as a standalone brand might not survive.
Since then, coverage from NotebookCheck, OC3D, HardwareZone, and TweakTown through 2025 and into 2026 has repeatedly used the combined term “RDNA 5/UDNA” to describe the same leaked dies, codenames, and specifications, including Navi 51, AT0, and AT2. That suggests either the branding decision hasn’t been finalized internally at AMD, or outlets are hedging by using both names until AMD confirms one. Either way, readers searching for “RDNA 5” and “UDNA” right now are largely finding coverage of the same underlying leaks.
Competitive Positioning: Where This Leaves Nvidia and Intel
AMD isn’t leaking RDNA 5 details in a vacuum. Nvidia currently dominates the discrete GPU market, and its RTX 50-series pricing has already drawn criticism this year after reported price hikes across the lineup. If the RDNA 5 flagship genuinely lands between an RTX 4090 and RTX 5080 in raw performance, as NotebookCheck’s sourcing suggests, AMD would be targeting the upper-midrange to lower-high-end segment rather than an outright halo part designed to beat Nvidia’s best card outright. That’s consistent with AMD’s RDNA 4 strategy, where the company skipped a true flagship entirely and focused the RX 9070 XT on value against Nvidia’s RTX 5070 Ti-class cards.
Intel, meanwhile, has been building out its own Arc GPU roadmap and continues to face delays on its data center AI silicon, leaving the consumer discrete GPU market largely a two-horse race between AMD and Nvidia heading into 2027. A genuinely new architecture from AMD, rather than another RDNA 3 refinement, would be the company’s clearest opportunity in years to close the performance-per-dollar gap that has kept Nvidia’s market share dominant.
Why a 2027 Product Is Making Headlines in 2026
The timing gap here is unusually wide, and it may be getting wider. Early estimates were considerably rosier than where the rumor mill sits today: Tom’s Hardware projected in September 2025 that AMD’s UDNA/RDNA 5 architecture would arrive sometime between late 2026 and early 2027, and industry speculation circulating on LinkedIn that same November pushed the date even earlier, floating a Q2 2026 release window. Neither held up. TweakTown’s June 2026 roadmap report had already pointed to a mid-2027 launch window for RDNA 5, but more recent reporting suggests that window is slipping further: Digital Foundry reported in June 2026 that the launch could slip to early 2028, roughly three years after RDNA 4’s March 2025 debut, while PC Games Hardware relayed that same month that new RDNA 5 Radeon cards aren’t expected before mid-2027 to 2028. Wccftech reported in June 2026 that partners are targeting Q3 2027 for RDNA 5 GPUs, with some partners expecting RDNA 5 as late as early 2028. VerdictBits reported in July 2026 that AMD’s own board and OEM partners now expect availability to land in mid-to-late 2027 rather than the sooner dates earlier rumors had suggested. TweakTown updated the picture again in September 2026, noting that leaker Kepler_L2 now sees the broad RDNA 5 gaming GPU rollout pushed to 2028, with only a single AT2 die, which that September 2026 reporting describes as a roughly 64-compute-unit RDNA 5 part, expected to arrive in 2027. Early leaks like this typically surface when engineering samples start circulating among AMD’s OEM and board partners, or when internal roadmap documents leak to trusted forum sources ahead of any official teaser. The gap between leak and launch also means specs, clock targets, and even branding can and often do shift substantially before a product ships. AMD’s own RDNA 4 generation saw its flagship plans change significantly between early leaks and the eventual RX 9070 XT launch.
That volatility is worth keeping in mind before treating any of these numbers as a purchasing decision input. Anyone shopping for a GPU today should base that decision on currently available RDNA 4 and RTX 50-series hardware, not on rumored 2027 clock targets that could easily move by hundreds of megahertz before AMD finalizes silicon.
Market and Investor Reaction
GPU architecture leaks rarely move AMD’s stock on their own, especially this far ahead of launch, but they do shape analyst sentiment heading into product cycles. AMD has increasingly framed its GPU roadmap around both gaming and AI data center silicon in the same breath, and any signal that RDNA 5 represents a genuine architectural leap, rather than an incremental refresh, feeds into the broader narrative that AMD can keep pace with Nvidia’s rapid GPU cadence across both markets. Investors watching AMD’s graphics division will likely treat this leak as a data point supporting that narrative rather than a standalone catalyst.
5 Predictions for RDNA 5’s Actual Launch
- Final retail boost clocks will likely land below the leaked 3.5 GHz ceiling for the flagship part, closer to the 3.0-3.2 GHz range once power and thermal targets are locked in, consistent with how the highest per-SKU numbers in these leaks tend to apply to the smallest, cheapest chips.
- AMD will probably keep marketing the product under some form of the RDNA or UDNA name rather than a fully rebranded architecture, given how much brand equity Radeon has built around the RDNA naming since 2019.
- Expect AMD to again skip an outright halo flagship aimed at beating Nvidia’s top card, instead targeting strong performance-per-dollar in the upper-midrange, mirroring the RDNA 4 strategy with the RX 9070 XT.
- The mid-2027 launch window cited by TweakTown’s roadmap leak will likely hold directionally, though specific quarters could slip if TSMC’s N3P capacity is constrained by AI chip demand from Nvidia and other buyers.
- Memory configurations will probably land closer to the more conservative 384-bit bus and lower VRAM figures reported by NotebookCheck and OC3D than the more aggressive 512-bit, 36 GB figures from HardwareZone and PCGuide, since AMD has historically prioritized cost control on mainstream and upper-midrange SKUs.
What to Watch Next
The next meaningful signal will likely come from AMD’s own conference appearances or investor calls rather than another forum leak. Companies typically start confirming architectural direction, if not full specs, roughly six to nine months ahead of a genuine launch, so an AMD confirmation is more likely to follow the reveal timing now rumored than to arrive this year. A report citing leaker Kepler_L2 says Computex 2027 would be the earliest plausible reveal, with a later-2027 announcement considered more likely. Until then, expect the 3.1-3.5 GHz clock range, the Navi 51 codename, and the RDNA 5/UDNA naming debate to keep resurfacing every time a new fragment of the same underlying leak gets reported by a different outlet.
Frequently Asked Questions
What clock speeds is AMD RDNA 5 rumored to hit?
Leaks reported by TechPowerUp and others point to a 3.1 GHz to 3.4 GHz target range for the RDNA 5 lineup, with a rumored ceiling of 3.5 GHz for an entry-level SKU. None of these figures are confirmed by AMD.
When will RDNA 5 GPUs launch?
Timing estimates have moved around. A June 2026 TweakTown report citing leaked roadmap data initially pointed to a mid-2027 launch window, but Digital Foundry reported that same month the launch could slip to early 2028, and TweakTown’s own September 2026 follow-up cites leaker Kepler_L2 describing the broad RDNA 5 rollout as pushed to 2028, with just one AT2 die expected in 2027. AMD has not announced an official release date.
Is RDNA 5 the same as UDNA?
It’s unclear. Early leaks from late 2024 suggested AMD might replace the RDNA 5 name entirely with UDNA, a unified architecture spanning gaming and data center GPUs. More recent 2025 and 2026 coverage uses “RDNA 5/UDNA” interchangeably to describe the same leaked specifications, so the final branding hasn’t been confirmed.
How much faster would RDNA 5 be than the RX 9070 XT?
Wccftech’s coverage puts the rumored 3.1-3.4 GHz clock range at nearly 14% higher than the RX 9070 XT’s 2.97 GHz factory boost clock. Separate leaked claims cited by TweakTown and PCGamesN suggest additional IPC gains of 5-10% and per-CU raster gains around 20%, though these figures come from unnamed leaker sources and are not confirmed.
What manufacturing process will RDNA 5 use?
Leaked roadmap reporting from TweakTown and Tech-Critter points to TSMC’s N3P node, a step up from the N4P process used for RDNA 4’s RX 9070 XT. This has not been confirmed by AMD or TSMC.
How many compute units will the RDNA 5 flagship have?
NotebookCheck and OC3D reporting from July 2025 points to 96 compute units for the flagship die, codenamed Navi 51, on a 384-bit memory bus. A separate HardwareZone report cites a wider 512-bit interface for the same die, and PCGuide’s leak references gaming CU counts as high as 154 in other spec tables, illustrating how much these numbers still conflict between sources.
Should I wait for RDNA 5 before buying a GPU?
Given that a report citing leaker Kepler_L2 puts Computex 2027 as the earliest plausible reveal, with a later-2027 announcement considered more likely, and the fact that none of these specs are confirmed, most buyers shopping today are better served evaluating currently available RDNA 4 and RTX 50-series cards rather than waiting on unconfirmed 2027 hardware.
Where do these RDNA 5 leaks come from?
The September 2026 clock-speed figures trace to a post from leaker Kepler_L2 on the AnandTech forums, first summarized by shattered.io and subsequently reported by TechPowerUp, Club386, VideoCardz, Wccftech, OC3D, and TwistedVoxel. Earlier 2025 leaks about CU counts, dies, and memory configurations were sourced separately by NotebookCheck, OC3D, HardwareZone, and PCGuide, citing Moore’s Law Is Dead and other industry sources.
