RDNA3 in the Works
Greymon shared that AIB partners are already testing AMD’s next-gen GPUs. If this is true, we may end up seeing benchmarks extremely soon.
— Greymon55 (@greymon55) October 19, 2022 Last month, AMD’s senior president, Sam Naffziger in a blog-post stated that the RDNA3 microarchitecture will allow for a 50% performance/watt improvement over RDNA2. Greymon claims that the reference variant, possibly for the RX 7900 XT has an ‘amazing TBP’ (Total Board Power). It could end up being lower than the RTX 4090’s 450W, as AMD has been much more power efficient than NVIDIA in the past few generations.
— Greymon55 (@greymon55) October 19, 2022
2x Faster Than Last-Gen
Taking things up a step further, Greymon mentioned that Navi31, the GPU powering the RX 7900 XT & the RX 7950 XT will be 2x faster (than Navi21?). This is a massive uplift from last-generation. Considering the fact that the RTX 3090 Ti and the 6950 XT are almost equal, the Navi31 powered 7950 XT could be on-par with the RTX 4090.
— Greymon55 (@greymon55) October 19, 2022 What’s better is that with the added raster performance improvements, AMD has also given special attention to the ray tracing department. Greymon claims that ray tracing is more than 2x faster (than Navi21). Now, whether these GPUs can compete against NVIDIA’s RT offerings is debatable. However, a marginal 5-10% performance difference may not mean much in the end if these GPUs offer higher efficiency.
— Greymon55 (@greymon55) October 19, 2022
2.4x Faster Compute Performance
As per angstronomics’s leaked specifications, Navi31 features 12288 Shading Units. That amounts up to around 96 CUs on the RDNA3 architecture (128 Shading Units / 1 CU). However, if we want a direct generation to generation comparison, RDNA2 and RDNA1 featured 64 Shading Units per each CU. Long story short, based on the legacy design Navi31 is packed with 196 CUs (if we assume 64 Shading Units / 1 CU). This has been done for the sake of comparison and to our surprise, Navi31 is equipped with 2.4x more Compute Units than Navi21. Not an apple fan highlights that the GA102 GPU powering the RTX 3090 Ti housed 84 SMs. The RTX 4090 with 128 SMs (1.5x more) is almost 1.6-1.7x faster than the RTX 3090 Ti. If we assume that these core count increases scale linearly for AMD as well, we may end up seeing a massive 2.5x performance leap generation on generation. A while back, Harukaze compliled a chart showing the ‘leaked’ TSE score for AMD’s kingpin Navi31 GPU. The RTX 4090 on average scores between 18,000–19,500 points in the same test. This puts Navi31 to be neck to neck against the 4090. As for the 4090 Ti, AMD may be planning for an X3D variant. Besides, the Ada Titan with the full AD102 die has been cancelled so the RTX 4090 Ti is possibly the highest end Lovelace variation we’ll see.
The Lack of Leaks
As stated above, we have not seen that many RDNA3 leaks thus far. One possibility could be that AMD is not giving any headroom to NVIDIA for making last minute adjustments before RDNA3’s announcement. If these GPUs are what they appear to be, NVIDIA could be forced to reduce prices even further. So far, we have the following leaked specifications for the RX 7000 lineup;
