AMD’s next-generation GPU architecture, expected to carry the RDNA 5 label, is taking clearer shape through a fresh wave of leaks. The desktop cards would be manufactured on TSMC’s advanced N3P process and reportedly carry the internal name AT, short for Alpha Triton. A top-tier configuration is said to house up to 192 Compute Units—double the figure found in the Radeon RX 7900 XTX—though industry sentiment strongly points to that full AT0 die being reserved for AI-focused products. Trimmed variants, likely designated AT1 and AT2, would instead form the backbone of the upcoming Radeon RX 10000 series.
Ambitious SKU stack emerges
New discussion points on NeoGAF have narrowed the focus to individual models and their expected specifications. At the entry point, a tentative Radeon RX 10050 XT is described with 24 Compute Units and raster performance approximately matching the GeForce RTX 3070. Memory configuration for this card could fall between 12 GB and 16 GB of LPDDR5X. One tier up, the projected RX 10060 XT doubles the compute count to 48 CUs, carries 16 GB to 24 GB of the same memory type, and aims for performance near the GeForce RTX 4080 Super.
An unusual memory choice for desktop
The presence of LPDDR5X on a desktop graphics card immediately raises questions. Low-power DRAM modules are engineered for tightly constrained mobile and embedded environments, while discrete GPUs traditionally have the thermal and electrical headroom to push higher power budgets in pursuit of peak speed. AMD has no history of using LPDDR memory on its mainstream desktop boards.
Several factors could make the approach viable. Modern compression technologies already help GPUs conserve bandwidth by reducing the volume of data shuttled between the chip and memory. Existing solutions include Delta Colour Compression, depth and framebuffer compression, and lossless memory compression. A further possible rationale is cost: LPDDR packages may allow AMD to simplify board design and lower manufacturing expenses on entry-level products. If RDNA 5 introduces more effective memory compression, then desktop cards could also serve as a proving ground before the technology trickles down to APUs and handheld processors, where LPDDR is ubiquitous and bandwidth remains the chief performance constraint.
Higher tiers bet on speed and ray tracing
The upper segment of the lineup reportedly changes course. The rumoured RX 10070 XT is said to feature 68 Compute Units paired with 16 GB to 24 GB of GDDR7 VRAM, with raster output comparable to the GeForce RTX 5080. The potential line-up skips the RX 10800 XT nameplate entirely. At the very top, a flagship RX 10090 XT is claimed to incorporate 96 Compute Units, between 24 GB and 36 GB of GDDR7 memory, and performance that approaches the RTX 5090.
More than raw frame rates, the leak emphasizes that RDNA 5 could deliver disproportionately large ray tracing gains compared to raster uplifts, potentially narrowing the long-standing gap with Nvidia. Such a shift would point to substantive architectural changes rather than a minor refresh. The generation has long been referred to in enthusiast circles as a possible “Zen moment” for Radeon, setting expectations high for a fundamental reset in competitiveness.
Open questions around power
Still missing from the picture is any concrete detail on power consumption. Even if the performance targets are met, the efficiency equation remains unknown. Should the flagship card operate within a thermal envelope similar to Nvidia’s RTX 5090 class, AMD may need to adopt the 12VHPWR power connector. Although revised connector versions have mitigated early problems, overall reliability continues to attract scrutiny from builders and reviewers alike.
Source: www.neogaf.com