Speculation about AMD’s upcoming graphics card architecture UDNA, which is being touted as the direct successor to RDNA 4, has been mounting for several weeks. While there have been rather vague hints about possible changes so far, a clear trend is emerging in the latest rumors: AMD could double the number of shaders per compute unit and thus initiate a noticeable restructuring of the architecture. At the heart of the discussion is the assumption that AMD could increase the current arrangement of 64 FP32 units per compute unit to 128 units. This adjustment would not only double the theoretical computing effort per CU, but would also significantly change the scaling in the large expansion stages. For the largest chip, which is referred to as AT0 in the rumors, up to 96 compute units are in the cards. Extrapolated, this would correspond to 12,288 FP32 ALUs, which would take AMD into dimensions similar to Nvidia’s top models.
The current speculations classify the upcoming generation into several chip variants, whereby it is particularly noticeable that an originally expected AT1 chip has apparently disappeared from the planning. This results in significant gaps in the memory interfaces, which can only be partially compensated for by configuration variants.
According to the rumors, the product range is currently as follows:
AT0 could appear as a flagship with up to 96 CU (12,288 shaders), equipped with eight shader engines, a 384- to 512-bit memory interface and 24 to 32 GiB of graphics memory.
AT2, with a maximum of 40 CU (5,120 shaders) and four shader engines, should be in the upper mid-range, with a bandwidth of 192 to 384 bits and memory variants between 12 and 24 GiB.
AT3 is expected to be smaller, with 24 CU (3,072 shaders), two shader engines, a 128- to 256-bit interface and 8 to 16 GiB.
AT4 forms the lower end with 12 CU (1,536 shaders) and a narrow memory connection of 128 to 164 bits, combined with 8 to 16 GiB.
The large gap between AT2 and AT0 is remarkable. Industry observers suspect that the unimplemented AT1 was originally intended for a 256-bit expansion. This would have allowed AMD to achieve a more even staggering of the memory interfaces. Another point of speculation concerns the memory technology used. While LPDDR5X could apparently be used in the smaller AT3 and AT4 models, the use of GDDR7 is being discussed for the larger variants. The move to GDDR7 would not only enable a significant increase in bandwidth, but would also open up unusual memory capacities thanks to 24 Gbit memory chips. This could give AMD the opportunity to work with unusual but cost-efficient configurations.
If you compare the specifications with Nvidia’s rumored plans for the RTX 5000 series, you can see parallels in the segmentation. Nvidia also has a considerable gap in its portfolio with the RTX 5090 and RTX 5080, which is characterized by different memory interfaces and expansion stages. AMD could therefore be deliberately focusing on a similar differentiation. However, it remains unclear whether AMD will offer the largest AT0 chip in its full configuration or whether deactivated units are planned for better chip yield. In the past, AMD has often relied on partial shutdowns for high-end models in order to be able to utilize more chips.
It is interesting that the various leakers and sources complement each other on some points. Names such as Zhangzhonghao, Kepler_L2 and MLID regularly appear in the discussions. While MLID still spoke of 192 compute units according to the previous counting scheme, this would exactly match the current rumors if the doubling of the shaders is taken into account. This at least creates a consistent picture, even if it is still not officially confirmed.
Conclusion
The potential architectural changes to UDNA would be a significant step for AMD. Doubling the number of shaders per compute unit could improve efficiency in certain scenarios while significantly increasing overall performance. At the same time, the rumors raise questions about the exact memory strategy, the segmentation of the chip variants and the final product selection. As AMD has not yet released any official information on this generation, all of this remains in the realm of speculation. What is clear, however, is that the rumored changes would be big enough to clearly differentiate UDNA from RDNA 4 and represent a real generational leap.
Source: PCGH
10 Antworten
Kommentar
Lade neue Kommentare
Mitglied
Urgestein
Mitglied
Urgestein
Mitglied
Urgestein
Urgestein
Urgestein
Urgestein
Urgestein
Alle Kommentare lesen unter igor´sLAB Community →