TBH yes I agree. I think we are reaching a clock rate max like we did with netburst. We can try and increase IPC even more, but I dont think that will be the primary shift.
Games will only scale with faster CPUs until the CPU is no longer the bottleneck. If you e.g. had a CPU 10x faster per core than current Raptor Lake CPUs, you wouldn't see much of a difference in most games, and most certainly not scale very far with the few games which do. But a few years down the line, you'll probably see a larger difference as game engines become more demanding of CPUs. Hopefully some of this is used for something
useful to improve the games, but unfortunately a lot of it will probably be bloat. This is already evident in those games which are very sensitive to large L3 caches, which is a symptom of bloated code.
With coding languages and game engines being more advanced then they were, I actually think we will start to see a shift in gaming whein the parallelism with either the engine, software stack or underlying technology APIs. honestly probably all of them.
For the most part, it's mostly secondary workloads we see more parallelized in games, like loading assets, audio processing or some other async tasks. The two most performance critical elements, the game simulation (which we used to call it "game loop" in the old days) and the main rendering thread will continue to dictate the performance of games. Smaller tasks may be delegated to smaller worker threads, but scaling this greatly increases the risk of delays which results in stutter, or worse glitches like we see in so many games now.
If multithreading is to provide significant gains in gaming performance in the future, there would have to be different kinds of changes than we've seen so far. As latency quickly adds up when trying to synchronize increasing number of threads, efforts to reduce latency or even "guarantee" deadlines would be required. Firstly a much faster OS scheduler, and probably some semi-"RT" like features so threads are undisturbed by other tasks. Secondly graphics drivers etc. would need to behave more like in a RT system, and thirdly possibly HW changes to streamline communication.
But while multithreading often gets the most attention, optimizing for ILP is much more important for performance scaling, whether it's for gaming or user interactive applications. For smaller work chunks which needs to be synchronized, multithreading can only get you so far before overhead or latency bottlenecks it, but modern CPUs are also increasingly superscalar, which means the relative performance gains for writing clean efficient code is larger than ever. And while CPU frontends are increasingly advanced, e.g. Meteor Lake improves branch misprediction recovery, the gains from saturating the pipeline is even greater. The bigger problem here is the software practises which are popular today, especially how OOP, abstraction and generalization are employed. It is remarkable how much having dense logic affects CPU performance. But at some point I would expect compilers and potentially ISAs to evolve in order to scale with wider CPU architectures, hopefully in a better way than Itanium.
So I guess Arrow Lake is going to be like Zen 5, Only slight improvements here and there and the main focus seems to be improving efficiency.
If anything, I'm hoping for more
consistent performance. Pushing clock speeds too far leads to very unstable clock speeds, and at least for some of us that may be more annoying than slightly lower but more consistent performance. That is at least my impression from comparing Raptor Lake(i5-13600K) to Comet Lake(i7-10700K) at work, purely anecdotal and subjective impression, even though Raptor Lake has clearly higher peak performance.
Hopefully this will reduce expectations and hype though about Arrow lake performance increases, Zen 5 had such a bad reception at launch due to overhype which led to massive disappointment at launch when people saw Zen 5 did not improve as much as they were expecting.
Overhype servers no one.
Unlike most, I'm not that disappointed with Zen 5, and I'm very curious to see how it performs in upcoming Threadripper models.