a leak has let us see the possible specifications of the Radeon RX 7900 XTa graphics card that should become the most powerful of the new generation of AMD, and that will arrive configured with a total of 6 minor chiplets that will integrate the L3 cachealso known as infinite cache, and the GPU, which in the end will not be divided into two interconnected chips.
I want to clarify what we have just commented so that you do not have doubts. When people started talking about the Radeon RX 7900 XT, and also about the Radeon RX 7800 XT, the possibility of AMD using an MCM design was being considered identical to what we had already seen in the Instinct series, that is, that it uses two interconnected GPUs to create a “super GPU”, in the same way that two CCD units are combined in the Ryzen 5000 to create a 16-core chip.
In the end, this forecast has not been fulfilled, since all the key elements of the Navi 31 GPU, which will give life to the Radeon RX 7900 XT, will be integrated into a single chip. Only the L3 cache will be spread across those six smaller chiplets. In the attached image I have indicated it so that you can see it more clearly.
Going back to the Radeon RX 7900 XT, this graphics card would be equipped with the Navi 31 core and manufactured in node de 5 nm. promises an improvement 50% yield per watt vs. RDNA2 and a totally restructured computing unit, which means that it will probably bring 128 shaders per compute unit (RDNA2 mounts 64 shaders per compute unit). This movement would be identical to what NVIDIA carried out with the GeForce RTX 30, a generation that doubled the number of shaders per SM unit.
Possible specifications of the Radeon RX 7900 XT
- Navi 31 graphics core with MCD (Multi-chiplet Die) design in 5nm node (L3 cache will come in 6nm).
- 12.288 shaders.
- GPU at 2.5 GHz and up to 3 GHz in turbo mode.
- 96 computing units. This reinforces the theory of the jump to 128 shaders per compute unit.
- 96 units for ray tracing acceleration, if AMD sticks to the same design we saw in RDNA2.
- Bus de 384 bits.
- 24 GB of GDDR6 memory at 18 GHz, which would leave us with a total bandwidth of 864 GB / s.
- 96 MB of infinite cache, 16 MB per chiplet.
- PCIe Gen5 interface.
We do not have data on the texturing unit settings or on the number of ROPs (raster units), but if AMD kept the distribution of 4 texturing units per CU the Radeon RX 7900 XT would have 384 texturing units. The number of raster units should be higher than the 128 that the Radeon RX 6900 XT mounts, a fact that I give you as a reference so that you have a more complete and precise idea of what we can expect from the Radeon RX 7900 XT.
The Radeon RX 7900 XT should max out at 75 TFLOPs of performance in FP32, a figure that would far exceed 23 TFLOPs from the Radeon RX 6900 XT. This data also reinforces that information that practically takes for granted the duplication of shaders per computing unit. The launch of the Radeon RX 7900 XT should take place between September and October of this year.