How important is CPU in games in 2026
>>743287697It's more about the architecture than clock speed.https://gamersnexus.net/pre-built-pc/valve-steam-machine-review-gpu-cpu-benchmarks-steamos-test-thermals-noise-and-price>CPU performance in Stellaris simulation time benchmarking had the Steam Machine at around the level of the R7 3700X or R5 3600.That's not very good. I used to have 3700X and it struggled in the first Kingdom Come when there were too many NPCs around.
>>743287791This. There's actually a physical limit to how fast a processor can go so newer processors improve the architecture to do things faster without needing to break Physics.
steam machine CPU is overkill considering the garbage GPU it's paired with
>>743287875Hey I have a 3060 and it's not garbage
That box is obsolete and unupgradeable Get this instead.
Well, depends on the game and the performance you expect! Some games aren't CPU-heavy at all, others are but this manifests in loading or turn-processing times and doesn't really impact moment-to-moment experience of clicking stuff, and then there are games (like Cities: Skylines, Kerbal Space Program and what not) that just are CPU-heavy.And like anon said, clock speed isn't that important. I'll try to explain why, with a bit of history (I'm not a domain expert, but I think I should be able to get the gist of it right). For parallelizable workloads core count matters too, but more importantly the bottleneck is in efficient use of that clock. The change over the decades has been granular, but if you think about 80286 CPU from the 80s for example, it basically does "what it says in the box": executing one machine code instruction at a time (some instructions, like multiplication, might take multiple clock cycles), and retrieving data from RAM, which is largely constrained by physics of physical distance, was relatively speaking speaking fast (a few clocks) at the time. Fast-forward to now and physics hasn't changed so retrieval from RAM is about as fast as it used to be, but clocks are 1000x higher, so what once took 5 clocks now takes 500. This has been addressed by having a memory hierarchy of caches of fast memory physically closer to the CPU, and the closest L1 cache has latency of maybe 5 cycles. But this requires the data to already be in the cache, so how well this mechanism works, and how much data the caches can store, could in a sense increase the speed by 100x (full utilization of cache vs everything is in RAM).
>>743289765The second difference is related: CPUs don't actually perform just one operation per cycle, but are "superscalar". Even with caches there's latency, so the trick is to build a pipeline of instructions, and the CPU can start working on the pipeline while waiting for data to arrive for the others. For this reason, some components (like ALU - arithmetic logical unit) are replicated and a single core could have as many as 8. So, in the naive picture (and reality in 80s and still mostly in 90s) a CPU does one addition per clock tick or one multiplication per 3 clock ticks, say, while in contemporary CPUs it could hypothetically do 8 additions in a single clock, or feed 8 multiplications to the pipeline, start working on them simultaneously, and get the answer for all 8 three cycles later (the latency is still 3 clocks, but the throughput is 1). And again, you can see here that 1) having those 8 ALUs; 2) actually utilizing them all the time, once more offers a potential 8x benefit compared to not utilizing them at all. It could even be worse actually, since some tasks depend on the previous ones, and modern CPUs *guess* (speculative execution - a lot of computing tasks are of form of doing an operation to 1st, 2nd, 3rd... element of an array, so if you notice that pattern then it's a good guess that you're going to be doing 4th-6th as well for example) and if they got the guess right then they keep on churning ahead but if they got it wrong then they have to roll back everything they've done and clean up and that takes a lot of cycles. So, the thing to look for nowadays is instructions per cycle (IPC).Here's a toy example, summing over an array the size of N:for (int i = 0; i < N; i++) sum += arr[i]; //Adding one element to the sum at a timevsfor (int i = 0; i < N; i += 4) s0 += a[i]; s1 += a[i+1]; s2 += a[i+2]; s3 += a[i+3]; //Keeping track of four "buckets", with the idea that a CPU would do those four sums simultaneouslyIt basically works.
>>743287949Would you buy a prebuilt with one for over $1000 in 2026 though?
I should be fine until the western collapse on 4k with my 5800x.
>>743287697The Cpu is faster than the PS5's yet overall weaker despite having a newer architecture. All games on Steam Machine run at lower resolution and Fps than on a base PS5.
>>743287791I think your problem is something else, even my i7 6800k didn't struggle with Kingdom Come Deliverance (unless you mean 2, I haven't tested that)
>>743287791>R5 3600That's my CPU from like 6 or 7 years ago, they really couldn't get something better?
>>743291745They could have gotten something better, there probably wasn't anything much better to choose from in the same power envelope without paying a lot more though.
When the AI bubble pops, do you think the market will be flooded with cheap used graphics cards and ram? Or will they just pile it all up and set it on fire
>>743292016RAM will get cheaper because the capacity is freed up.. Server GPUs are basically useless.
>>743291127Back then I had 2070 Super and 16 Gb of RAM, a 1080p monitor, and I obviously used an SSD. I got decent FPS in the woods, but it dropped to 40 in Rattay. So no, it was definitely a CPU bottleneck.
>>743287697not at allall the work is done on the GPU nowadaysif you have a 5090, you can just go with Pentuim III and it will still work perfectly fine
>>743292195Found an old screenshot of the game, 54 FPS in Rattay with maxed out GPU usage on the Steam Deck. There's no way that's a CPU bottleneck on something like a 3700X if even Steam Deck can do almost 60. That CPU is way stronger than a Steam Deck CPU. KCD also utilizes many threads, it loads them all up. 8 cores 16 threads is really good for it.
>>743292392Couldn't get 5090 to fit in my PCI slot
>>743293217There are PCI to PCIe adapters.
>>743292392this is bait
