The recent conjecture regarding AMD’s forthcoming RDNA 5/UDNA architecture suggests the potential for up to 96 compute units in gaming GPUs. The buzz around AMD’s new GPU generation is gaining pace. Four chip designs have reportedly been leaked, which are believed to be aimed at the next lineup of graphics cards. Notably, the count of shader units is on the rise once again following the absence of a high-end model with the RX 9000. Nonetheless, the leaked information does not imply any direct competition with Nvidia, at least not in name.
## Four chips in four sizes
Details have emerged from Kepler_L2, who is typically well-versed in this area, in the [Anandtech forum](https://forums.anandtech.com/threads/rdna-5-udna-cdna-next-speculation.2624468/page-40#post-41496872) elaborated on the four chip designs in greater detail. Internally, at least, the variants are referred to as “AT0,” “AT2,” “AT3,” and “AT4.”
– AT0 is the largest chip, with all subsequent dies smaller in a descending order.
– If you’re curious about the absence of AT1, this design has reportedly been scrapped, as noted by YouTuber “Moore’s Law is Dead” recently.
The premium AT0 model thus features 96 compute units (CUs). Following that is the AT2 design with 40 CUs for the performance sector, then AT3 with 24 CUs for mainstream uses, and AT4 with 16 CUs set as an entry-level option. All chips will be produced based on the UDNA architecture.
## Strategic architecture change
AMD revealed last year that the RDNA architecture’s end was approaching. It will be succeeded by the “Unified DNA” strategy, the new UDNA generation, which merges gaming and data center architecture under one framework. This integration aims to lower development costs while simultaneously enabling manufacturing economies of scale, as evident in the previously mentioned gradation. As is customary, UDNA is also organized hierarchically.
The AT0 chip comprises a total of eight shader arrays, each containing two shader engines. Each shader engine includes six compute units, resulting in the aforementioned 96 CUs in the case of the AT0 die.
**The only variation here is that the AT2 chip supposedly possesses only five CUs per shader engine, which accounts for the projected 40 CUs.**
## The rough calculation for the number of shaders
If, as before, 64 shader units are implemented per CU, we arrive at a total of 6,144 shader units for the anticipated top model. That would still represent 50 percent more than the RX 9070 XT, yet it remains significantly behind the RTX 5090 with 21,760 shader units. This is largely attributed to the “density” with which Nvidia organizes its version of the CUs: there are 128 shader units per component in this case.
Even in an ideal scenario where AMD indeed utilizes 128 shader units per CU, one would only achieve 12,288 computing units – which would still surpass the level of an RTX 5080.
Incidentally, based on the same calculation with 64 shader units per CU, the other chips yield the following quantities:
– **AT2:** 2,560 shader units
– **AT3:** 1,536 shader units
– **AT4:** 768 shader units
## AMD divides all dies into several categories
In this context, another strategic shift is on the horizon, as the leaker “Mareeswj” points out on the X platform. With the phase-out of the previously mentioned AT1 chip, which creates a significant void in the proposed UDNA portfolio, the era of “standalone gaming chips” appears to be closing, at least for the moment. AT0 will not only serve Radeon graphics cards but also potential “prosumer” cards such as a Radeon VII.
– AT2 is intended for the next generation of traditional consoles. Both the Xbox Next and the PlayStation 6 will again employ AMD technology.
– The much smaller AT3 and AT4, conversely, are aimed at tablets and handheld devices.
As you can see, the current situation is not particularly clear. It will take some time before we receive any concrete information: UDNA production is not anticipated to commence until the second quarter of 2026. Therefore, we are unlikely to see actual graphics cards until the fall of 2026 at the earliest.