AMD RDNA 5 AT2 Leak: 70 CUs and Up to 3.4 GHz?
New AMD RDNA 5 rumors point to a GPU die reportedly known as AT2, with hardware insider Kepler_L2 claiming the chip could feature 70 Compute Units (CUs) and target clock speeds between 3.1 GHz and 3.4 GHz.
If accurate, those specifications would represent a clear increase over AMD’s current RDNA 4 flagship, the Radeon RX 9070 XT. Simple theoretical calculations suggest the AT2 die could deliver roughly 14% to 25% more raw math throughput, depending on the final clock speed.
However, these figures remain unconfirmed, and theoretical compute gains should not be interpreted as direct gaming-performance predictions.
🔍 Reported AMD AT2 Specifications #
According to the leak reported by OC3D, AMD’s RDNA 5-based AT2 die is expected to feature:
| Specification | RDNA 5 AT2 | Radeon RX 9070 XT |
|---|---|---|
| Compute Units | 70 CUs | 64 CUs |
| Target/Boost Clock | 3.1–3.4 GHz | Up to 2.97 GHz |
| Architecture | RDNA 5 | RDNA 4 |
| Status | Unconfirmed leak | Released product |
The reported 70 CUs would represent about a 9.4% increase over the RX 9070 XT’s 64 CUs.
At the upper end of the leaked clock range, 3.4 GHz would also be approximately 14.5% higher than the RX 9070 XT’s 2.97 GHz maximum boost clock.
📊 Theoretical Throughput Could Rise by Up to 25% #
A basic way to estimate the change in theoretical shader throughput is to compare the product of CU count and clock frequency.
Using the RX 9070 XT as the baseline:
- RX 9070 XT: 64 × 2.97 GHz
- AT2 at 3.1 GHz: 70 × 3.1 GHz
- AT2 at 3.4 GHz: 70 × 3.4 GHz
This produces an estimated raw throughput increase of approximately:
| AT2 Clock | Estimated Raw Throughput Gain |
|---|---|
| 3.1 GHz | ~14.4% |
| 3.4 GHz | ~25.1% |
The calculation illustrates the potential hardware-level improvement, but it does not account for architectural differences between RDNA 4 and RDNA 5.
Why the Calculation Is Only Theoretical #
CU count and clock speed are not enough to predict final GPU performance.
The actual performance of a future RDNA 5 product will also depend on factors such as:
- Instructions-per-cycle improvements
- Cache architecture
- Memory bandwidth
- Power limits
- Front-end and scheduling efficiency
- Driver optimization
- Ray-tracing performance
- AI acceleration
- Game-engine optimization
Consequently, a 25% increase in this simplified compute calculation would not automatically translate into 25% higher gaming performance.
⚡ Could RDNA 5 Challenge the RTX 5080? #
Some unconfirmed RDNA 5 rumors suggest AMD could make substantial improvements to ray tracing and AI acceleration.
Those claims have led to speculation that an AT2-based product could compete above the performance level of NVIDIA’s GeForce RTX 5080. However, the leaked CU count and clock speeds alone are insufficient to establish such a performance position.
AMD has not officially confirmed:
- AT2 as an RDNA 5 product designation
- 70 CUs
- 3.1–3.4 GHz clock targets
- The final architecture’s IPC improvements
- Ray-tracing performance
- AI acceleration performance
- Final product positioning
Any comparison with the RTX 5080 therefore remains speculative until independent hardware benchmarks become available.
🎮 AT2 Is Also Linked to Xbox Project Helix #
The AT2 rumor has another potential connection: Microsoft’s next-generation Xbox, currently known as Project Helix.
The leak reportedly claims that the console could use a customized version of the AT2 design with 68 enabled CUs, potentially leaving two CUs disabled for yield-management purposes.
If accurate, the configuration would look like this:
| GPU Design | Reported CU Count |
|---|---|
| Desktop AT2 | 70 CUs |
| Project Helix custom variant | 68 CUs |
This connection remains unconfirmed, however, and should be separated from Microsoft’s official Project Helix announcements.
What Microsoft Has Confirmed #
Microsoft has confirmed that Project Helix is being developed with AMD around a custom AMD SoC. The platform is designed to take advantage of next-generation DirectX and FSR technologies, with Microsoft also highlighting major improvements in ray-tracing performance.
Microsoft has not publicly confirmed that the console SoC uses an AT2 die or disclosed an exact CU count or GPU clock speed.
What Remains Unconfirmed #
The following details currently belong to the rumor category:
- AT2 as the Project Helix GPU design
- 68 enabled CUs
- The relationship between the desktop and console versions
- Console GPU clock speeds
- Memory configuration
- Final GPU architecture
- Final retail specifications
A custom console SoC could also differ substantially from a desktop GPU even if both designs share architectural technology.
🗓️ Reported RDNA 5 Launch Timeline #
Kepler_L2 reportedly expects AT2-based products to arrive in 2027, potentially making the design AMD’s primary new desktop GPU introduction for that year.
Other RDNA 5 designs aimed at additional market segments are rumored to follow in 2028.
These dates are not part of an official AMD roadmap.
| Item | Reported Timeline | Official Status |
|---|---|---|
| AT2-based GPU | 2027 | Unconfirmed |
| Broader RDNA 5 lineup | 2028 | Unconfirmed |
| Official RDNA 5 roadmap | Not announced | — |
Release schedules for unreleased GPUs can change significantly as silicon validation, manufacturing capacity, product segmentation, and market conditions evolve.
🧠 What the AT2 Leak Could Mean for RDNA 5 #
If the reported specifications are accurate, AT2 would represent a substantial increase in raw hardware resources over the RX 9070 XT, combining 70 CUs with a target clock range of 3.1–3.4 GHz.
The simplified calculations suggest roughly 14% to 25% more raw compute throughput, before accounting for any architectural improvements introduced by RDNA 5.
The more important unknowns are therefore not limited to CU count and clock speed. RDNA 5’s architectural efficiency, ray-tracing capabilities, AI acceleration, memory subsystem, power targets, and final product configuration will determine how much of that theoretical advantage reaches real-world applications and games.
For now, AT2 remains a leaked design rather than an officially announced AMD product, while its reported connection to Project Helix and the 2027 launch timeline also require further confirmation.