If his data is accurate (and given the recent Nvidia hack, it could very well be), Ada will be a huge upgrade over Ampere, an RTX 30-series, and a flagship GPU for better choice. If you have not heard of this data or not, then take these results with a bit of salt.
J’ai made a simple graph for this site.twitter.com/zilwXgi0vaMarch 1, 2022
The leak shows that Nvidia won’t change its nomenclature for the Ada generation, while offering two-letter prefix and three-digit number as the Ampere generation. The AD102 denotes the largest GPU, likely an RTX 3090 or a Titan-class card, with the AD103 following the second most powerful die (possibly a potential RTX 4080). The 104-106 will follow suit with high-speed dies (ie RTX 4070 and RTX 4060), and the AD107 will take advantage of Nvidia’s Ada GPUs for the rise-level market (ie a RTX 4050). Note that the codenames suggest that Nvidia will use the Ada and not the rumored Lovelace codename, so we’re going to refer to future GPUs now. The number of SM in Ada has changed. The latest AD102 is a whopping 144 SMs in one das with the scales to ask. In comparison to Ampere’s GA102, only 84 SM is present, so that’s a 71% increase in SM count. That should be followed by GPU cores, RT cores, etc. This is the biggest leap we have ever seen in the entire generation. Since Nvidia has the same number of CUDA cores on Ada, we could go with the flagship card for 18432 CUDA cores. A full GA102 chip is Nvidia’s new RTX 3090 TI for the full solution. Of course we’ll see smaller variants that use partially harvested AD102 chips, while 144 SMs can be the maximum, so we won’t be surprised to see 100% of the SMs disabled in some graphics card models. That’s not the most important number of SMs in the other chips, but the number’s still very large. AD103 will have the same 84 SM as GA102 with the same 84 SMs. This is a 40% increase from GA103. The AD104 has 60 SMs as GA103, or 25% more SMs than GA104. AD106 is closer to GA106, with 36 SMs getting a 20% uplift. Finally, the AD107 will presumably have just 24 SMs, again with the respectable 20% jump in the SM count to GA107.
If these leaks and rumors prove true, we can expect that the future RTX 4090 and RTX 4080 will be loaded with impressive performance improvements over the current RTX 30 series. There is definitely a bigger jump than Ampere, at least in some respects. The example for the RTX 3080 had 68 SMs, though there were plenty of other changes. The above does not account for any further performance improvement from the Ada architecture itself, and so could other benefits. Recently, it’s been rumored that Ada will get it back to TSMC on its latest 5nm phone call. This alone should improve efficiency and the transistor count over the ampere and also unlock higher clock speeds. With the addition of the new 1262-pin power connectors, they are also developed and manufactured for future PCIe 5.0 graphics cards. With a power output of 600W in one plug, Nvidia can be built to boost Ada GPUs. Ada may be the first 5.0-compatible graphics solution, and the increase in PCIe bandwidth might not be too much – otherwise it would have significantly affected performance. What we do not know is how much Nvidia plans to change the basic building blocks of Ada. For example, Turing had 64 FP32 and 64 INT32 cores per SM, and then managed to run multiple data types simultaneously. Ampere altered the performance of the INT32 cores so that they become INT32 or FP32 cores. Ampere also has ray tracing for three tidal and two trelal cores. Ada will likely use fourth generation Tensor cores and threth generation RT cores. What’s it going to mean? We don’t have the exact details, but the Ada will certainly deliver much more performance than the currently used Ampere GPUs. There can be many additional CUDA, Tensor and / or RT cores per SM, or the internal pipelines that can only be revamped so that it increases throughput. Memory is a really good choice when it comes to GPU performance. It could also play an even bigger role, with higher frame rates, and the number of SMs Ada has. The GDDR6 + and GDDR7 are already Samsung’s roadmap – besides the GDDR6X, and Nvidia will probably use one or both of these new standards if they are ready for Ada production. Since most of the cores are built, you need more memory to feed them all. Generally speaking, Nvidia has increased its performance on its fastest GPUs by around 30% with previous architectures. However, with the changes in process nodes and massively increased core counts, plus a much higher power limit, it does not mean to expect even greater improvements from Ada. Will the RTX 4090 (or whatever it’s going to be called) deliver twice the performance of the RTX 3090? It’s ambitious, but definitely not reachable. Seven5% more cores would be more than one clockspeed and / or a more efficient architecture. This year we’ll find out more, as Ada is expected to launch in September.